Information processing system

Abstract

The application provides an information processing system. An information processing system, characterized by comprising: a processor; wherein the processor is configured to: receive a commercial property characteristic and a user attribute as input; generate a prompt text for indicating that a generative artificial intelligence model generates an asset text based on the input information; automatically generate an asset text based on the prompt text; apply the generated asset text to an advertisement; and control the display of the advertisement.

Description

Technical Field

[0001] The technology disclosed herein relates to an information processing system. Background Technology

[0002] Japanese Patent Application Publication No. 2022-180282 discloses a method for controlling a role-based chatbot executed by at least one processor. The method includes the following steps: receiving a user's speech; adding the user's speech to a prompt word, the prompt word containing instruction statements associated with an explanation of the chatbot's role; encoding the prompt word; and inputting the encoded prompt word into a language model to generate a chatbot response to the user's speech.

[0003] In the current online advertising process, the generation and optimization of advertising copy (asset text) mainly rely on manual writing or simple rule splicing, making it difficult to personalize it in a timely and accurate manner according to the characteristics of the product and user attributes. This results in a low degree of matching between advertising content and user needs, and low click-through rate and conversion rate.

[0004] On the other hand, although some systems attempt to personalize advertisements to a certain extent based on user behavior, they usually only make adjustments to the selection of advertising materials or the display position. They do not make full use of generative artificial intelligence models to dynamically generate and optimize the advertising copy itself, and they are unable to efficiently build high-quality and diversified asset texts on a large scale and in multiple scenarios.

[0005] Furthermore, with the rapid changes in user behavior and interests, existing technologies struggle to update ad copy in real time based on the latest search and browsing behaviors. As a result, ad content lags behind changes in users' actual interests, reducing the attractiveness and effectiveness of ad placement.

[0006] Therefore, it is necessary to provide a system that can automatically generate and dynamically optimize asset text using generative artificial intelligence models, and can also be updated in real time based on user attributes and behaviors, in order to improve the matching degree between advertisements and user needs and enhance the efficiency and effectiveness of advertising. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides an information processing system comprising a processor configured to: receive product characteristics and user attributes as input; generate prompt text based on the input information to instruct a generative artificial intelligence model to generate asset text; automatically generate asset text based on the prompt text; apply the generated asset text to an advertisement; and control the display of the advertisement. Through this configuration, the system can automatically generate personalized advertising asset text based on specific product information and individual user characteristics, thereby reducing manual writing costs and improving the relevance of copywriting to user needs.

[0008] Preferably, the processor is further configured to optimize the asset text based on the user's behavioral history. By analyzing the user's past search behavior, browsing behavior, click behavior, and purchase behavior, the processor can make targeted modifications or rewrites to the generated asset text, making the asset text more aligned with the user's long-term interests and preferences, thereby improving the relevance and conversion rate of the advertisement.

[0009] Furthermore, the processor is configured to update the asset text in real time based on the user's attributes and behaviors. That is, when the user's attribute information changes, or when new user behaviors such as searching or browsing are detected, the processor can regenerate the prompt text and invoke a generative artificial intelligence model to generate new asset text, or adjust the existing asset text in real time and immediately apply the updated asset text to the advertisement for display. Through this real-time update mechanism, the advertising content can dynamically reflect the user's current interest state and changes in needs, thereby significantly improving the attractiveness of the advertisement and the user experience.

[0010] Through the above technical solution, this invention combines generative artificial intelligence models with user profiles and user behavior data to achieve automatic generation of advertising asset text, optimization based on behavioral history, and real-time updates. This overcomes the problems of low efficiency, insufficient personalization, and untimely response to changes in user interests in existing technologies, thereby achieving the technical objective of improving the efficiency and effectiveness of advertising.

[0011] A "system" refers to a collection of devices consisting of one or more hardware components, software modules, and their cooperating functional units, used to perform a series of processing operations such as receiving data, generating asset text, applying asset text to advertisements, and controlling the display of advertisements.

[0012] "Processor" means a hardware processing unit or combination thereof for executing program instructions, including but not limited to a central processing unit (CPU), a graphics processing unit (GPU), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), and other computing units that can be configured by software to implement the functions of this invention.

[0013] "Product characteristics" refers to attribute information related to the goods or services to be advertised, including but not limited to product category, price range, brand information, target audience, functional features, promotional information, and other characteristic data that can be used to describe the goods or services.

[0014] "User attributes" refer to static or relatively stable characteristic information related to the target users of advertising, including but not limited to the user's age, gender, geographical location, interests, occupation, income level, device type, and other user characteristics obtained through user registration or data analysis.

[0015] "Generative AI models" refer to AI models that can automatically generate text content based on input prompts, including but not limited to deep learning-based natural language generation models, large-scale pre-trained language models, and other generative models that can output asset text through training.

[0016] "Prompt text" refers to informational text generated by the processor based on asset characteristics and user attributes, used to provide instructions to the generative artificial intelligence model. This text is used to constrain or guide the generative artificial intelligence model to output asset text that conforms to the expected content and style.

[0017] "Asset text" refers to the displayable text content used in advertising, including but not limited to ad headlines, ad descriptions, selling point copy, promotional copy, and other text fragments related to the ad creative. This text is automatically generated by a generative artificial intelligence model based on prompt text or processed from its generation results.

[0018] "Advertisement" refers to an information display unit for promoting goods or services, consisting of one or more display elements, including but not limited to the text portion of text advertisements, graphic advertisements, and video advertisements, as well as comprehensive display information containing asset text, images, links, and other content.

[0019] "Display advertising" refers to the process and result of presenting advertising content on the interface of a user terminal device, including but not limited to outputting advertising information in the form of a graphical user interface on web pages, mobile applications, smart TVs or other devices with display functions.

[0020] "User behavior history" refers to dynamic behavioral data related to users recorded within a certain time frame, including but not limited to users' search history, browsing history, click history, collection history, purchase history, and other interactive behavioral information that can reflect users' interests and preferences.

[0021] "Optimizing asset text" refers to the process of adjusting, rewriting, supplementing, or simplifying the generated asset text based on user behavior history and other relevant data in order to improve the matching degree between the asset text and user interests, needs, and delivery goals.

[0022] "Real-time asset text update" refers to the process where, when user attributes or user behavior change, the processor regenerates or modifies the asset text under conditions close to the actual time of occurrence, and applies the updated asset text to the advertisement display immediately or within a very short delay. Attached Figure Description

[0023] Figure 1 This is a conceptual diagram illustrating an example of the configuration of the data processing system according to the first embodiment.

[0024] Figure 2 This is a conceptual diagram illustrating an example of the main functions of the data processing apparatus and smart device according to the first embodiment.

[0025] Figure 3 This is a conceptual diagram illustrating an example of the configuration of the data processing system according to the second embodiment.

[0026] Figure 4 This is a conceptual diagram illustrating an example of the main functions of the data processing device and smart glasses according to the second embodiment.

[0027] Figure 5 This is a conceptual diagram illustrating an example of the configuration of the data processing system according to the third embodiment.

[0028] Figure 6 This is a conceptual diagram illustrating an example of the main functions of the data processing device and head-mounted terminal according to the third embodiment.

[0029] Figure 7 This is a conceptual diagram illustrating an example of the configuration of the data processing system according to the fourth embodiment.

[0030] Figure 8 This is a conceptual diagram illustrating an example of the main functions of the data processing device and robot according to the fourth embodiment.

[0031] Figure 9 This represents an emotion map that maps multiple emotions.

[0032] Figure 10 This represents an emotion map that maps multiple emotions.

[0033] Figure 11 This is a sequence diagram illustrating the processing flow of the data processing system of the first embodiment.

[0034] Figure 12 This is a sequence diagram illustrating the processing flow of the data processing system in Application Example 1.

[0035] Figure 13 This is a sequence diagram illustrating the processing flow of the data processing system of the second embodiment.

[0036] Figure 14 This is a sequence diagram illustrating the processing flow of the data processing system in Application Example 2. Detailed Implementation

[0037] Hereinafter, an example of an implementation of the system according to the present disclosure will be described with reference to the accompanying drawings.

[0038] First, let me explain the terminology used in the following instructions.

[0039] In the following embodiments, the processor (hereinafter referred to as "processor") with reference numerals may be a single computing device or a combination of multiple computing devices. Furthermore, the processor may be a single computing device or a combination of multiple computing devices. Examples of computing devices include CPU (Central Processing Unit), GPU (Graphics Processing Unit), GPGPU (General-Purpose computing on Graphics Processing Units), APU (Accelerated Processing Unit), etc.

[0040] In the following embodiments, RAM (Random Access Memory), as indicated in the figures, is a memory that temporarily stores information and is used as working memory by the processor.

[0041] In the following embodiments, the memory, as indicated by the reference numerals, is one or more non-volatile storage devices that store various programs and parameters. Examples of non-volatile storage devices include flash memory (SSD (Solid State Drive)), disks (e.g., hard disks), or magnetic tapes.

[0042] In the following embodiments, the communication I / F (Interface) with reference numerals is an interface that includes a communication processor and an antenna, etc. The communication I / F is responsible for communication between multiple computers. As an example of a communication specification applicable to the communication I / F, wireless communication specifications such as 5G (5th Generation Mobile Communication System), Wi-Fi (wireless fidelity) (registered trademark), or Bluetooth (registered trademark) can be listed.

[0043] In the following embodiments, "A and / or B" is synonymous with "at least one of A and B". That is, "A and / or B" means that it can be only A, only B, or a combination of A and B. Furthermore, in this specification, when "and / or" connects to express more than three items, the same interpretation as "A and / or B" applies.

[0044] First Implementation Method Figure 1 An example of the configuration of the data processing system 10 according to the first embodiment is shown.

[0045] like Figure 1 As shown, the data processing system 10 includes a data processing device 12 and an intelligent device 14. A server can be cited as an example of the data processing device 12.

[0046] The data processing apparatus 12 includes a computer 22, a database 24, and a communication I / F 26. The computer 22 is an example of a "computer" as understood in this disclosure. The computer 22 includes a processor 28, RAM 30, and memory 32. The processor 28, RAM 30, and memory 32 are connected to a bus 34. Furthermore, the database 24 and the communication I / F 26 are also connected to the bus 34. The communication I / F 26 is connected to a network 54. Examples of the network 54 include a WAN (Wide Area Network) and / or a LAN (Local Area Network).

[0047] The smart device 14 includes a computer 36, a receiving device 38, an output device 40, a camera 42, and a communication I / F 44. The computer 36 includes a processor 46, RAM 48, and memory 50. The processor 46, RAM 48, and memory 50 are connected to a bus 52. In addition, the receiving device 38, output device 40, camera 42, and communication I / F 44 are also connected to the bus 52.

[0048] The receiving device 38 includes a touchscreen 38A and a microphone 38B, and receives user input. The touchscreen 38A receives user input via touch by detecting contact with an indicator (e.g., a pen or finger). The microphone 38B receives user input via sound by detecting the user's voice. The control unit 46A in the processor 46 sends data representing the user input received by the touchscreen 38A and microphone 38B to the data processing device 12. In the data processing device 12, the specific processing unit 290 acquires the data representing the user input.

[0049] The output device 40 includes a display 40A and a speaker 40B, and presents data to the user 20 by outputting data in a form perceptible to the user 20 (e.g., sound and / or text). The display 40A displays visual information such as text and images according to instructions from the processor 46. The speaker 40B outputs sound according to instructions from the processor 46. The camera 42 is a small digital camera equipped with an optical system such as a lens, aperture, and shutter, and an imaging element such as a CMOS (Complementary Metal-Oxide-Semiconductor) image sensor or a CCD (Charge Coupled Device) image sensor.

[0050] Communication I / F44 is connected to network 54. Communication I / F44 and 26 are responsible for sending and receiving various information between processor 46 and processor 28 via network 54.

[0051] Figure 2 The diagram shows an example of the main functions of the data processing device 12 and the smart device 14.

[0052] like Figure 2 As shown, in the data processing apparatus 12, specific processing is performed by the processor 28. A specific processing program 56 is stored in the memory 32. The specific processing program 56 is an example of a "program" as understood in this disclosure. The processor 28 reads the specific processing program 56 from the memory 32 and executes the read specific processing program 56 on the RAM 30. Specific processing is implemented by the processor 28 operating as a specific processing unit 290 according to the specific processing program 56 executed on the RAM 30.

[0053] The memory 32 stores a data generation model 58 and an emotion-specific model 59. The data generation model 58 and the emotion-specific model 59 are used by the specific processing unit 290. The specific processing unit 290 can use the emotion-specific model 59 to infer the user's emotions and perform specific processing based on the user's emotions. In the emotion inference function (emotion-specific function) using the emotion-specific model 59, various inferences and predictions related to the user's emotions are performed, including inferences and predictions of the user's emotions, but this is not limited to this example. Furthermore, emotion inference and prediction may also include, for example, emotion analysis (parsing).

[0054] In the smart device 14, the processor 46 performs the acceptance output processing. The memory 50 stores the acceptance output program 60. The acceptance output program 60 is used in conjunction with the data processing system 10 and the specific processing program 56. The processor 46 reads the acceptance output program 60 from the memory 50 and executes the read acceptance output program 60 on the RAM 48. The acceptance output processing is implemented by the processor 46 acting as the control unit 46A according to the acceptance output program 60 executed on the RAM 48. Furthermore, the smart device 14 has the same data generation model and emotion-specific model as the data generation model 58 and the emotion-specific model 59, and these models can also be used to perform the same processing as the specific processing unit 290. The acceptance output processing is implemented by the processor 46 acting as the control unit 46A according to the acceptance output program 60 executed on the RAM 48.

[0055] Alternatively, other devices besides the data processing device 12 may also have the data generation model 58. For example, a server device (e.g., a generation server) may have the data generation model 58. In this case, the data processing device 12 obtains the processing results (prediction results, etc.) using the data generation model 58 by communicating with the server device that has the data generation model 58. Furthermore, the data processing device 12 may be a server device or a user-held terminal device (e.g., a mobile phone, robot, home appliance, etc.). Next, an example of the processing of the data processing system 10 of the first embodiment will be described.

[0056] Example 1 The flow of a specific process in Example 1 will be described. Each part of the system described below is implemented by the data processing device 12 and the smart device 14. Furthermore, the data processing device 12 is referred to as the "server," and the smart device 14 is referred to as the "terminal."

[0057] With the rapid growth of online advertising scale and user numbers, the process of generating advertising copy increasingly relies on computer systems for automation. However, in existing technologies, servers typically input product information and user attributes directly into generative AI models by concatenating strings, resulting in the model generating advertising text in a "black box" manner. This presents the following technical problems: First, the server side lacks a structured processing and standardization mechanism for input data, leading to coarse-grained generation of input prompts. It is difficult to finely distinguish based on multi-dimensional product characteristic information and user attribute information uploaded from different terminals, thus limiting the quality and controllability of the model output. Second, the server does not modularize the prompt construction process, failing to automatically adjust the model input based on explicit generation conditions (such as style conditions, word count conditions, and output quantity conditions), resulting in inconsistent generation outcomes under the same computing resources. First, the system struggles to reliably meet the format and constraint requirements of different advertising scenarios. Second, the server typically lacks a systematic post-processing workflow for the asset text output by the generative AI model, failing to automatically determine and correct prohibited words, word counts, and formatting at the computer level. This necessitates significant manual intervention, reducing the overall throughput of the advertising generation system. Third, in utilizing historical and real-time user behavior information, the server primarily uses it for statistical analysis, without forming a closed-loop control mechanism that feeds back to prompts and generation conditions. This prevents the system from dynamically optimizing the input of the generative AI model based on historical and real-time data, thus limiting the effectiveness of computer technology in improving advertising responsiveness and personalization through algorithmic improvements.

[0058] Therefore, there is a need for a mechanism on the server side that improves the advertising copy generation process driven by generative artificial intelligence models through structured data processing, configurable prompt statement construction, automated post-processing, and dynamic optimization based on historical and real-time behavioral data. This would improve the overall computer technology performance of the advertising generation system in terms of accuracy, controllability, real-time performance, and computational efficiency without relying on additional hardware resources.

[0059] The specific processing performed by the specific processing unit 290 of the data processing apparatus 12 in Embodiment 1 is achieved by the following means.

[0060] In this invention, the server includes: a device for receiving product characteristic information and user attribute information from a terminal and storing the information as structured data; a device for performing information extraction and normalization processing on the structured data to form element information containing product information elements and target user information elements; a device for automatically assembling prompt statements for a generative artificial intelligence model according to a preset template based on the element information and generation condition information including style conditions, word count conditions, and output quantity conditions; a device for sending the prompt statements and generation parameter information to an information processing device running the generative artificial intelligence model via a network, and obtaining asset text generated by natural language generation processing from the information processing device; and a device for performing violations on the asset text. The system includes a device for post-processing such as prohibited word detection, character count determination, and format sorting to generate advertising asset data; a device for embedding the advertising asset data into an advertising delivery data structure, generating advertising display data for display terminals, and sending it to the terminals; a device for associating and storing product characteristic information, user attribute information, and user behavior history information as historical information, and dynamically updating prompt statements and generation condition information based on the historical information to optimize the output results of the generative artificial intelligence model; and a device for continuously parsing user behavior information from the terminals, changing element information or generation condition information in real time according to the parsing results, and triggering the regeneration processing of the generative artificial intelligence model with the updated prompt statements to update the asset text in real time. This allows for the replacement of traditional simple string concatenation and static calling methods with modular structured processing, configurable prompt statement construction, and automatic post-processing on the server side, achieving fine control over the input and output of the generative artificial intelligence model. Furthermore, the system continuously optimizes the generation process through a feedback mechanism of historical and real-time behavior data, significantly improving the computer performance of the automatic advertising copy generation system in terms of output quality, personalization, real-time response capability, and processing throughput per unit of computing resources.

[0061] "Product characteristic information" refers to attribute information related to the product or service to be promoted, including but not limited to category, price range, functional features, usage scenarios, target audience, and other structured or unstructured data used to describe the characteristics of the product or service.

[0062] "User attribute information" refers to static or relatively stable characteristic information related to the target user, including but not limited to age, gender, region, interests, occupation, income level, and other attribute data used to create a user profile.

[0063] "Structured data" refers to data that is organized and stored according to a predetermined data model or field format, including data that exists in the form of tables, key-value pairs, records, etc., and can be directly indexed and processed by programs.

[0064] "Information extraction processing" refers to the process of automatically analyzing and parsing the received raw data to identify and extract key information elements of predefined types.

[0065] "Standardization processing" refers to the process of performing operations such as format unification, encoding conversion, case unification, synonym unification, and illegal character cleanup on data content in order to improve data consistency and processability in subsequent processing stages.

[0066] "Product information elements" refer to the basic information units used to describe the core characteristics of a product after being extracted and standardized from product characteristic information, including but not limited to product category, price level, functional tags, and usage scenario tags.

[0067] "Target user information elements" refer to the basic information units used to describe the characteristics of target users after being extracted and standardized from user attribute information, including but not limited to age group, gender category, interest tags, and behavioral preference tags.

[0068] “Element information” refers to a structured set of information formed by combining product information elements and target user information elements, which is used to construct and generate input.

[0069] "Generation condition information" refers to the set of control parameters used to constrain and guide generative artificial intelligence models in producing text content, including but not limited to stylistic conditions, word count conditions, output quantity conditions, tone requirements, and content style requirements.

[0070] "Style requirements" refer to the constraints or requirements imposed on the generated text in terms of language style, expression, formality, and emotional tone.

[0071] "Word count condition" refers to the constraint parameter set on the length range or maximum number of words of the generated text, which is used to control the length of the output text.

[0072] "Output quantity condition" refers to the parameter setting for the expected number of text entries or candidate results in a single generation process.

[0073] "Generative artificial intelligence models" refer to artificial intelligence models that are based on machine learning and deep learning technologies and can automatically generate natural language text or other content based on input prompts.

[0074] "Prompt statements" refer to instructional or guiding texts input into generative artificial intelligence models, which contain product information elements, target user information elements, and generation condition information to guide the model to generate asset text according to expected requirements.

[0075] "Generation parameter information" refers to the model operation parameters set when calling a generative artificial intelligence model, including but not limited to temperature value, maximum output length, sampling strategy, generation mode, and other control parameters related to the model's inference behavior.

[0076] "Information processing device" refers to a computing device or collection of computing resources that has computing power and is capable of performing generative artificial intelligence model inference processing, including but not limited to servers, cloud computing platforms or acceleration hardware devices.

[0077] "Natural Language Generation Processing" refers to the computational process by which a generative artificial intelligence model automatically generates text content that conforms to the grammar and semantics of natural language based on input prompts and generation parameter information.

[0078] "Asset text" refers to natural language text content generated by generative artificial intelligence models based on prompts and used as candidates for advertising copy.

[0079] "Post-processing" refers to the process of detecting prohibited words, verifying word count, formatting, standardizing punctuation, and performing necessary text trimming or minor revisions after obtaining the asset text.

[0080] "Prohibited word determination" refers to the process of determining whether asset text contains predefined sensitive words or prohibited terms through keyword matching, rule detection, or other algorithms.

[0081] "Word count determination" refers to the process of counting the actual number of words in the asset text and comparing it with the preset word count condition to determine whether the length requirement is met.

[0082] "Formatting" refers to the standardization and normalization of line breaks, indentation, punctuation, numbering methods, etc. of asset text to meet the format requirements of display or downstream systems.

[0083] "Advertising asset data" refers to a collection of post-processed, qualified asset text and related structured descriptive information, which is used to embed in the data structure for advertising placement.

[0084] "Data structure for ad delivery" refers to the data organization format designed for managing and transmitting ad content in an ad delivery system, including but not limited to structures containing title fields, description fields, creative identifier fields, and targeting parameter fields.

[0085] "Ad display data" refers to advertising content data that is transformed from advertising placement data structures and is suitable for direct presentation on display terminals, including text, links, and related display control information.

[0086] "User behavior history information" refers to historical behavioral data of user interactions with content or systems recorded over a period of time, including but not limited to click records, browsing records, dwell time, conversion behavior, and interaction logs related to advertisements.

[0087] "Historical information" refers to a data set obtained by associating product characteristic information, user attribute information, and user behavior history information, which is used to provide a basis for generation and optimization in subsequent processing.

[0088] "User behavior information" refers to data collected from the terminal in real time or near real time that reflects the user's current or recent interaction, including but not limited to page browsing, click operations, scrolling behavior, input behavior, and interactive feedback.

[0089] "Continuous parsing" refers to a processing mode in which the server continuously receives and parses constantly updated user behavior information, and updates the analysis results in a streaming or periodic manner.

[0090] "Regeneration processing" refers to the process of calling the generative artificial intelligence model again to generate a new round of text for the same advertising scenario after updating the prompt statement or generating condition information.

[0091] The embodiments of this invention will combine the roles of the server, terminal, and user in the system, and will specifically describe the system structure, data structure, algorithm flow, and internal mechanism of the generative artificial intelligence model, so that those skilled in the art can implement this invention based on this description.

[0092] In one implementation, the server runs on a computing device with network communication and parallel computing capabilities, such as a network server or cloud computing node deployed with a multi-core general-purpose processor. The server operating system can be a UNIX-like system, such as a Linux server operating system. The server deploys network service middleware, such as an HTTP server and a backend application framework; the server is also connected to an external generative artificial intelligence model inference service, which runs on computing resources configured with graphics processing units, such as a cluster of graphics processing units supporting large-scale matrix operations. In this invention, the server performs a series of data processing and computational operations through the aforementioned hardware and software structure, including structured data processing, prompt statement construction, generative artificial intelligence model invocation, and asset text post-processing.

[0093] In one implementation, a terminal is a user-held smart terminal device, such as a mobile terminal or tablet terminal equipped with a touchscreen and a wireless network interface. The terminal's operating system can be a mobile terminal operating system or another graphical interface operating system. The terminal runs an application for interactive ad generation or a browser-based front-end page, which communicates bidirectionally with the server through the network stack provided by the operating system.

[0094] In this invention, users input product characteristic information and user attribute information using a graphical interface provided by the terminal. The terminal collects the results of operations such as text input, option selection, and tag checking through input method software and graphical controls, converting the user input into internal data objects. In this invention, the terminal does not execute complex generation algorithms, but mainly plays the role of data collection, data display, and interactive control, thereby concentrating most of the computational load on the server.

[0095] After receiving product characteristic information and user attribute information uploaded by the terminal, the server converts this information into structured data. In one implementation, the server uses a key-value structure, record structure, or relational record structure to represent the structured data. For example, the server can use records containing multiple fields, each including a field name, field type, and field value, to represent product category, price level, tag set, target age group, gender category, interest tag set, etc. The server uses a data validation module to check the field types and value ranges, discarding or replacing values ​​that do not conform to the rules, thereby ensuring the consistency of the data received by subsequent calculation modules.

[0096] The server performs information extraction and normalization on structured data. In one implementation, information extraction utilizes keyword matching, rule template matching, and simple feature classification algorithms to identify core semantic fragments from the raw text. For example, based on a predefined product category dictionary, the server maps "women's tops" and "hoodies" to unified category identifiers; based on a predefined price range, it maps "300-800 yuan" to the "mid-price range" label. Normalization, in one implementation, includes operations such as character encoding standardization, stop symbol removal, case uniformity, synonym unification, and redundant word removal. The server completes these steps using a string processing library and rule tables. Through the above extraction and normalization, the server generates product information elements and target user information elements, such as "mid-price range women's tops," "women in their 20s," and "outdoor activity preferences."

[0097] The server combines product information elements and target user information elements into element information, and generates generation condition information based on this. In one implementation, the generation condition information includes text style conditions, word count conditions, and output quantity conditions. For example, the server can set the word count condition to "no more than 60 characters per piece" based on the advertising platform's field restrictions; set the text style condition to "young, lighthearted, and visually appealing" based on the advertising scenario; and set the output quantity condition to "output 3 different pieces of copy" based on the need for multi-version testing. This generation condition information is stored in a structured format, such as key-value pairs representing "tone=young", "max_length=60", "count=3", etc.

[0098] The server constructs prompts for a generative artificial intelligence model using element information and generation condition information. In one implementation, the server uses a template engine module that maintains various text templates to map structured data into natural language descriptions. The server selects an appropriate template based on the advertising scenario, inserts product information elements and target user information elements into placeholders within the template, and then adds stylistic and word count requirements to form a complete prompt. For example, the server can generate the following prompt: Please generate 3 Chinese advertising copy based on the following conditions: - Product Information: Mid-priced women's top, suitable for spring and summer outdoor, camping, and travel activities. - Target users: Women in their 20s who enjoy the outdoors, camping, and travel. Require: - Each piece of copy should not exceed 60 characters. - The tone is youthful, relaxed, and vivid. - Avoid using English words and emojis. In the aforementioned process, the server specifically designs the sentence structure, paragraph order, and presentation of constraints for the prompts to improve the controllability and stability of the generative AI model. Compared to simply concatenating keywords, this structured construction method makes it easier for the model to identify task requirements and constraints during reasoning, thereby improving the matching accuracy and consistency of the generated results.

[0099] The server sends the constructed prompt and generation parameter information over the network to the information processing device running the generative AI model. In one implementation, the generation parameter information includes temperature parameters, maximum generation length, and sampling strategies. The temperature parameter controls output diversity; the server can set a moderately high temperature value based on the advertising scenario to obtain candidate texts that are differentiated yet within constraints. The maximum generation length parameter limits the number of characters the model outputs; the server calculates a reasonable token limit based on the character count and explicitly sets it in the request. The sampling strategy can be a top-k or top-p sampling strategy based on probability distributions, used to avoid repetitive and monotonous expressions while maintaining mainstream semantics.

[0100] In one implementation, the generative artificial intelligence model is a deep neural network based on a multi-layer transformer architecture. This network includes multiple self-attention sublayers and feedforward sublayers. It encodes the input prompts into a high-dimensional vector sequence through word embedding layers and calculates the internal correlations of the sequence using a multi-head self-attention mechanism. During the training phase, the model utilizes a large-scale corpus, calculates the error between the generated and target sequences using a cross-entropy loss function, and updates the weight parameters using a stochastic gradient descent-like optimization algorithm. Data augmentation and regularization methods, such as random masking, Dropout, and weight decay, can be used during training to improve the model's generalization ability. In the inference phase, this invention utilizes the trained model parameters to perform forward propagation calculations on the prompts, obtaining the conditional probability distribution of the next symbol at each time step, and generates asset text through a sampling strategy.

[0101] The server receives asset text output by a generative artificial intelligence model from an information processing device. In one implementation, the asset text may contain multiple candidate advertising copy, separated by line breaks or sequence numbers. The server first splits the asset text into a list of individual copy entries using a format parsing module, and then performs post-processing on each copy. Post-processing includes prohibited word detection, word count determination, and format correction. The prohibited word detection module uses a pre-built sensitive word list and matching algorithm to detect whether the text contains terms that are not allowed in advertisements; the word count determination module counts the number of words in the text and compares it with the word count criteria; the format correction module standardizes sentence-ending punctuation, numbering formats, spaces, and line breaks. For copy that does not conform to the rules, the server can discard, trim, or initiate a regeneration request if necessary. By centrally managing the post-processing logic on the server side, this invention embeds text quality control into the computational process, reducing the burden on the terminal side and the human side.

[0102] After post-processing, the server generates advertising asset data. In one implementation, this data is represented as structured records, containing fields such as copy text, associated product identifiers, associated user group identifiers, and quality scores. The quality score can be calculated based on a combination of post-processing results, historical click feedback, and model confidence levels. The server embeds this advertising asset data into an advertising delivery data structure, which is adapted to the specific advertising delivery system's field requirements, such as title, description, creative identifiers, and delivery parameters. Based on this, the server generates advertising display data for the display terminal and sends it to the terminal or other display devices via the network.

[0103] After receiving the ad display data, the terminal uses its local rendering module to combine the ad copy with images, videos, and other materials to present to the user. The terminal can adjust line breaks and font size based on screen size and interface style to improve readability. User behavior information, such as clicks, dwell time, and scrolling patterns, is collected by the terminal and sent back to the server. The server continuously analyzes this user behavior information, associating it with corresponding ad asset data, product characteristic information, and user attribute information to form historical user behavior information. Based on this historical information, the server adjusts its prompt message construction strategy and generation conditions, such as increasing the weight of high-conversion copy styles and reducing the probability of low-performing styles, thereby achieving adaptive optimization at the algorithm level.

[0104] In another implementation, the server triggers in-line regeneration based on real-time user behavior information. When the server detects a significant decrease in the interaction rate of a certain user group with existing ad copy during a specific time period, it can automatically change element information or generation condition information, such as strengthening promotional attributes, highlighting contextual descriptions, reconstructing prompts, and calling the generative AI model again. Through this closed-loop control, the server not only automatically generates ad copy but also adjusts its generation strategy through continuous feedback, enabling the system to gradually improve output quality and relevance over long-term operation.

[0105] In this invention, the server employs the aforementioned structured data processing, templated prompt statement construction, model parameter control, and post-processing mechanisms. This transforms the invocation of generative artificial intelligence models from a simple linear automation of human copywriting into a fine-grained control of model input and output through specialized internal computer data structures and algorithmic processes. This control yields the following technical effects: First, by pre-extracting element information and generation condition information on the server side, prompt statements are more explicit and structurally sound, reducing repeated model attempts and invalid outputs, thereby lowering network calls and computational resource consumption, and improving overall processing throughput. Second, through a unified post-processing and quality control mechanism, the probability of unqualified text appearing on the terminal side is significantly reduced, improving the stability and availability of end-to-end system output. Third, through a dynamic adjustment mechanism based on historical information and real-time behavioral data, the server can improve the "effectiveness ratio" of model calls without increasing hardware resources, i.e., obtaining more results that meet business and format requirements with the same number of inferences, substantially improving the efficiency of computational resource utilization.

[0106] In another implementation, the server can employ different generative AI model structures. For example, it can add a conditional encoding module for control vectors to the basic transformer architecture to explicitly encode control signals such as stylistic conditions and word count conditions. The server can also use multi-task training methods to simultaneously minimize language modeling loss and style classification loss during model training, allowing the model to better adhere to style constraints in the prompt statements during generation. Furthermore, the server can perform embedding space retrieval on the prompt statements before calling the model, calculating the similarity between the current prompt statement and historically high-performing prompt statements, and adjusting the current generation condition information by referencing the generation results of similar historical samples, thereby further improving the quality and stability of the generated results.

[0107] In another implementation, the server can deploy a lightweight rule engine internally to dynamically update the list of prohibited words and format rules based on advertising platform specifications and laws and regulations. The rule engine manages rule priorities in a tree-like or directed acyclic graph manner. During the post-processing stage, the server applies rules sequentially according to priority. This non-simple linear scanning method can reduce redundant checks while ensuring accuracy, thus improving the computational efficiency of the post-processing stage.

[0108] In summary, this invention, by introducing specific data structures, a transformer-based generative artificial intelligence model control method, and a dynamic optimization mechanism based on user behavior feedback on the server, enables the automatic advertising copy generation system to achieve data management optimization, improved generation accuracy, enhanced computational efficiency, and reasonable control of communication load at the computer technology level. It represents an improvement on the internal information processing and model reasoning process of computers, rather than simply an automated replacement of human copywriting tasks.

[0109] use Figure 11 The processing flow is explained.

[0110] Step 1: Users input product characteristic information and user attribute information using the terminal.

[0111] In the graphical interface of the terminal, users can input or select product characteristic information such as product category, price range, target audience, usage scenario, and promotional information, as well as user attribute information such as age group, gender, interests, and region, through components such as text input boxes, drop-down menus, and checkboxes.

[0112] The terminal takes the raw character sequence and option values ​​entered by the user as input, uses local data structures (such as key-value pairs or objects) to collect and validate fields, sets incomplete fields to null values, filters out illegal characters, and serializes the results into structured request data (such as a set of key-value pairs). This structured request data is then sent to the server for subsequent use.

[0113] Step 2: The terminal sends structured request data to the server.

[0114] The terminal takes the structured request data generated in step 1 as input, calls the network communication interface provided by the operating system, constructs a request message via HTTP or HTTPS protocol, adds identification information to the request header, and embeds the structured request data in the request body. The terminal then performs a network transmission operation, sending the request message as output to the interface address specified by the server via the network.

[0115] Step 3: The server receives and parses the structured request data sent by the terminal.

[0116] The server takes the request message sent by the terminal as input, parses the HTTP / HTTPS protocol layer data through the network service framework, and extracts the structured request data from the request body. The server performs JSON parsing or form parsing on this data, converting the text-based key-value pairs into an internal data structure, such as a dictionary or a list of records. During parsing, the server performs existence checks on required fields, type checks on field types (e.g., price ranges are strings, age ranges are interval strings), and fills missing or abnormal data with default values ​​or marks them as errors. The parsed and validated internal data structure is then used as output for subsequent extraction and normalization processing.

[0117] Step 4: The server performs information extraction processing on the parsed data.

[0118] The server takes the parsed and validated internal data structure as input, calls the information extraction module, and extracts product information elements and target user information elements from the original fields according to a pre-built dictionary and rules. Through string matching, rule mapping, and simple classification logic, the server maps "women's tops" to a standard category tag, "300-800 yuan" to "mid-price range," and "likes camping and travel" to the interest tag "outdoor activities." The server writes the extraction results into a new set of fields, forming product information elements and target user information elements containing standardized tags and semantic classifications, and outputs this set of elements.

[0119] Step 5: The server performs normalization processing on the feature information.

[0120] The server takes the set of elements output from step 4 as input and performs standardization operations on each element, including character encoding unification, removal of redundant spaces and punctuation, synonym unification (e.g., unifying "outdoor" and "outdoor activities" into the label "outdoor activities"), and pruning of lengthy descriptions. The server uses string processing functions and a standardization rule table to uniformly standardize expressions from different sources. The standardized element information is written back into a structured data object, forming standardized element information, which is used as output for subsequent construction of generation conditions and prompt statements.

[0121] Step 6: The server generates the necessary conditions.

[0122] The server takes standardized element information and system configuration or advertising platform constraints as input, and calculates applicable style conditions, word count conditions, and output quantity conditions through a rule engine. For example, the server selects "youthful, relaxed, and vivid" as the style condition based on the placement, sets the word count condition "each piece of copy cannot exceed 60 characters" based on platform field restrictions, and sets the output quantity condition "output 3 pieces of copy" based on testing requirements. The server organizes the above conditions into a structured generated condition information object (containing fields such as tone, max_length, and count), which is then used as output by the prompt statement construction module.

[0123] Step 7: The server constructs prompt statements based on feature information and generation condition information.

[0124] The server takes standardized element information and generation condition information as input, calls the template construction module to select a suitable text template from the template library, and inserts product information elements and target user information elements into the reserved placeholder positions in the template. Simultaneously, the server converts the style conditions, word count conditions, and output quantity conditions into natural language constraint descriptions and appends them to the end of the template. The server uses string concatenation and formatting functions to combine these fragments into a complete prompt statement, and neatens line breaks and punctuation. Finally, the server outputs a prompt statement text that can be directly sent to the generative artificial intelligence model, for example: Please generate 3 Chinese advertising copy based on the following conditions: - Product Information: Mid-priced women's top, suitable for spring and summer outdoor, camping, and travel activities. - Target users: Women in their 20s who enjoy the outdoors, camping, and travel. Require: - Each piece of copy should not exceed 60 characters. - The tone is youthful, relaxed, and vivid. - Avoid using English words and emojis. Step 8: The server constructs a request to invoke a generative artificial intelligence model.

[0125] The server takes the prompt message output in step 7 and the generation condition information output in step 6 as input, and combines them with the system's preset model call parameters (such as model type, temperature value, maximum number of tokens, and sampling strategy) to construct a model request object. The server places the prompt message in the "User Message" field, maps the generation conditions to the model parameter field, and adds model identifiers and authentication information. Through serialization, the model request object is converted into a network request message as output, ready to be sent to the information processing device running the generative artificial intelligence model.

[0126] Step 9: The server sends a request to the information processing device where the generative artificial intelligence model resides and receives the asset text.

[0127] The server takes the network request message constructed in step 8 as input, calls the network client library, and sends the message to the remote information processing device via HTTPS or other secure protocols. Upon receiving the message, the remote information processing device embeds and encodes the prompt statement, performs matrix multiplication, self-attention calculation, and feedforward calculation using a multi-layer transformer structure, and generates a text sequence token-by-token according to a set sampling strategy. This generated sequence is encapsulated as a response message and returned as model output. After receiving the response message, the server parses out the asset text containing advertising copy candidates and outputs it as plain text or a list of multiple texts for subsequent post-processing.

[0128] Step 10: The server performs prohibited word checks and character count checks on the asset text.

[0129] The server takes the asset text list output from step 9 as input and calls the prohibited word detection module and the word count module one by one. The prohibited word detection module uses a sensitive word dictionary and matching algorithm to scan each text for sensitive words and mark whether it violates the rules; the word count module compares the number of characters in the text with a pre-set word count condition to determine whether it exceeds the length limit. The server appends the detection result of each text to the corresponding document record, distinguishing between qualified and unqualified texts; for obviously unqualified texts, the server can mark them as discarded or ready for secondary generation. After the above judgment, the set of asset text records with quality labels is passed as output to the formatting module.

[0130] Step 11: The server formats and organizes qualified asset texts to generate advertising asset data.

[0131] The server takes the asset text records deemed qualified or acceptable in step 10 as input and calls the formatting module to uniformly adjust the text numbering, sentence-ending punctuation, spaces, and line breaks. The server can remove unnecessary numbering prefixes, delete redundant blank lines, and ensure that each piece of text is independent and complete. After formatting, the server adds product identifier, user group identifier, and internal quality score fields to each piece of text, constructing an advertising asset data record. All advertising asset data records are organized into a structured set, which is used as output by the advertising delivery data structure generation module.

[0132] Step 12: The server generates the data structure for ad delivery and constructs the ad display data.

[0133] The server takes the set of advertising asset data as input and maps each asset data point to an advertising delivery data structure according to the target advertising platform or front-end display format definition. This includes filling in fields such as title, description, creative identifier, and targeting parameters. The server also adjusts the field combinations based on the terminal type; for example, it prioritizes short text fields for mobile devices and uses longer description fields for desktop devices. After encoding, the set of advertising delivery data structures generated by the server forms an advertising display data packet as output, ready to be sent to the terminal or other display terminals.

[0134] Step 13: The server sends advertising display data to the terminal.

[0135] The server takes the ad display data packet generated in step 12 as input, encapsulates it into a response message or push message through the network service framework, and sends it to the specified terminal address using HTTP / HTTPS or push protocols. The server appends a version number or request identifier to the message to facilitate result mapping and cache management between the terminal and the server. The terminal receives the response at the network layer and passes the ad display data packet as output to the local rendering module.

[0136] Step 14: The terminal renders and displays the advertisement, and the user generates behavioral feedback.

[0137] The terminal takes the advertising display data sent by the server as input, calls the interface rendering engine, fills the text content into predefined interface components, and displays the combined image or video materials with the text on the screen. During user browsing, the terminal uses an event listening mechanism to record behavioral data such as clicks, scrolling, and dwell time, and organizes this data into behavioral logs. At set time intervals or when specific events are triggered, the terminal sends these behavioral logs back to the server via the network for subsequent historical information updates and generation strategy optimization.

[0138] Step 15: The server updates historical information and adjusts the generation strategy based on user behavior information.

[0139] The server takes user behavior logs sent from the terminal as input, associates the behavior data with corresponding advertising asset data, product characteristic information, and user attribute information to form historical user behavior information, and writes it to the historical information storage module. Through statistical analysis and simple modeling, the server calculates the relationship between specific copywriting styles, specific prompt statement structures, and actual performance (such as click-through rate and conversion rate), and updates the weight configuration or template selection strategy for generating conditional information accordingly. The updated strategy parameters are stored as new configuration data and used as output for the next round of prompt statement construction and model invocation, thereby achieving adaptive optimization based on data feedback.

[0140] Application Example 1 The process flow corresponding to the specific processing in Use Case 1 will be described below. The various parts of the system described below are implemented by the data processing device 12 and the intelligent device 14. Furthermore, the data processing device 12 is referred to as the "server" and the intelligent device 14 is referred to as the "terminal".

[0141] With the rapid growth of information and product volume in online services, traditional advertising systems generally employ strategies based on fixed rules or simple scoring models, matching pre-written ad templates with a limited number of user attributes. These systems typically suffer from the following technical problems: First, server-side processing of user and product attribute data often remains at the level of simple filtering and coarse-grained tag matching, failing to perform fine-grained modeling of high-dimensional behavioral data and rich product features, resulting in insufficient ad relevance. Second, the process of generating ad content is disconnected from the user behavior data processing process in the system architecture. Servers often simply assemble text based on static templates, failing to deeply integrate user profiles, product profiles, and prompts from generative AI models within a unified program flow. Third, existing systems rely heavily on manually written generic prompts when calling generative AI models, lacking consideration of input conditions, output conditions, and text length. The algorithmic generation mechanism of the constraints leads to uncontrollable generation results, inconsistent styles, and low matching degree with target users and target products; fourth, the post-processing of generated copy in traditional systems is often relatively simple, lacking systematic content filtering capabilities such as prohibited expression detection, information consistency verification, and automatic length adjustment, which easily produces text containing sensitive words or inconsistent with the actual information of products, resulting in technical defects in compliance and user experience; fifth, although user behavior data such as browsing, clicking, and closing on the terminal is collected, it is rarely fed back to the update link between prompts and generated copy in real time, lacking an online optimization mechanism driven by behavior feedback, making it difficult for the advertising system to form an adaptive closed loop at the model call level.

[0142] Therefore, it is necessary to propose a new system architecture and data processing method that enables the server to: perform unified preprocessing and feature extraction of user attribute information and product attribute information; automatically generate prompt statements corresponding to user profiles and product profiles; on this basis, controllably call generative artificial intelligence models to generate asset text; perform multi-dimensional content filtering and correction on the generated results; and update user profiles and interest levels in real time based on user behavior information returned from the terminal, thereby improving the advertising content generation and delivery process at the computer technology level, and enhancing the overall data processing efficiency, model call controllability, and online personalization capabilities of the system.

[0143] The specific processing performed by the specific processing unit 290 of the data processing apparatus 12 in Application Example 1 is achieved by the following means.

[0144] In this invention, the server includes a unit for acquiring and preprocessing attribute information related to products and attribute information related to users; a unit for extracting feature quantities from the attribute information and calculating the relevance between multiple product candidates and the target user, and selecting target product candidates for generating prompt statements accordingly; a unit for generating natural language expressions containing user profiles and product profiles based on preprocessed user attribute information and product attribute information, and simultaneously specifying input conditions, output conditions, and word count constraints for the generative artificial intelligence model; and a unit for inputting the prompt statements into the generative artificial intelligence model and automatically generating data composed of advertising text elements through deep learning inference. The system comprises the following units: a unit for generating asset text; a unit for performing prohibited expression detection, information consistency verification, and length adjustment on the generated asset text to filter and correct its content; a unit for embedding the filtered and corrected asset text into advertising display data and sending the advertising display data to the terminal device via a communication unit; a unit for parsing the advertising display data on the terminal device and displaying the advertisement containing the asset text in real time on the user's display device; and a unit for receiving user behavior information regarding browsing, selection, or operation of the advertisement from the terminal device, updating user attribute information and interest level information accordingly, and feeding the update results back to subsequent prompt statements and asset text generation processes. This allows for the formation of a closed-loop technical process within the same computer system, encompassing structured processing of user and product data, automatic construction of prompt statements, controllable invocation of generative artificial intelligence models, multi-level content filtering of generated text, and online profile updates based on behavioral feedback. This enhances the server's ability to process high-dimensional data and its real-time personalized generation capabilities, achieving computer technology improvements in the advertising content generation and delivery process, and significantly improving advertising relevance and system operating efficiency.

[0145] A "system" refers to a collection of devices that consist of multiple hardware and software components that collaborate with each other through communication to perform functions such as user data processing, product data processing, prompt generation, generative artificial intelligence model invocation, asset text generation, and advertising display control.

[0146] A "server" refers to a computing device in a system that performs functions such as data storage, data processing, model inference, and communication control. It may include one or more processors, memory, and communication interfaces, and is used to execute the program processing corresponding to each unit described in this invention.

[0147] "Terminal equipment" refers to information processing devices used to display advertisements to users and collect user behavior information, including computing devices with display devices and network communication functions, such as mobile terminals, tablet terminals or other electronic terminals with display and interactive functions.

[0148] "Product-related attribute information" refers to a set of data used to characterize the features of a target product, including but not limited to product category, price or price range, promotional information, target audience tags, functional descriptions, and other product characteristic data that can be used for ad generation and matching.

[0149] "User-related attribute information" refers to a set of data used to characterize the features of a target user, including but not limited to the user's age, gender, geographic location, interest tags, browsing history, search history, purchase history, and preference and value level information inferred from user behavior.

[0150] "Preprocessing" refers to the process of cleaning, completing, standardizing, encoding, or format conversion of raw data to transform it into structured data that is easier for subsequent feature extraction, model input, and logical processing.

[0151] "Features" refer to quantitative data extracted from product-related attribute information and user-related attribute information, used to represent the characteristics of products and users in vector space, including numerical features, category coding features, embedding vectors, etc.

[0152] "Product candidates" refers to a set of one or more products that can be used as targets for ad generation, selected from multiple available products based on their relevance to the target user.

[0153] "Relevance" refers to a quantitative indicator used to measure the degree of matching between a specific product candidate and a target user. It can be obtained through feature vector similarity calculation, scoring models, or other algorithms, and is used to rank and filter product candidates.

[0154] "User profile" refers to a structured or semi-structured description of a user based on user-related attribute information and user behavior information, which is formed by modeling the user's age group, gender, interests, spending power, behavior habits, etc. It can be used to generate prompts and personalized advertising content.

[0155] "Product profile" refers to a structured or semi-structured description of a product based on its attributes, such as category positioning, price level, target audience, main selling points, and usage scenarios. It can be used to generate prompts and personalized advertising content.

[0156] "Prompt statements" refer to natural language text used as input to generative artificial intelligence models. They include at least natural language expressions of user profiles and product profiles, and explicitly specify the model's input conditions, output conditions, language type, style, and word count constraints to guide the model in generating asset text that meets expectations.

[0157] "Generative artificial intelligence models" refer to machine learning models that can automatically generate natural language text based on input prompts. They typically employ deep learning structures, learning from large-scale corpora during the training phase and generating semantically relevant output text based on the input text during the inference phase.

[0158] "Asset text" refers to text elements that constitute advertising content, automatically generated by generative artificial intelligence models based on prompts, including but not limited to headlines, selling point descriptions, promotional information, and calls to action.

[0159] "Prohibited Expression Detection" refers to the process of identifying and marking prohibited words, sensitive terms, or expressions that do not meet predetermined compliance standards in generated asset text using keyword lists, rules, or classification models.

[0160] "Information consistency verification" refers to the process of comparing key information such as price, discount, and product attributes in the generated asset text with the actual product attribute information stored in the system in order to detect and correct inconsistencies or errors.

[0161] "Length adjustment" refers to the process of controlling the number of words or characters in the generated asset text, including deleting or truncating parts that exceed the length limit, or triggering regeneration when necessary to meet the preset word count constraints.

[0162] "Ad display data" refers to a structured collection of data used to display advertisements on terminal devices, including at least asset text, media resource identifiers related to the product, jump link information, and control information related to the ad display location or style.

[0163] "User behavior information" refers to the records of user interactions such as browsing, clicking, swiping, closing, and dwell time during the display of advertisements on terminal devices. These records are used to reflect the user's level of interest in and response to advertisements and products.

[0164] "Interest level information" refers to quantitative data inferred from user behavior information about the strength of a user's preference for a certain type of product, a certain type of advertising content, or a certain type of theme. It can be used to update user profiles and adjust subsequent advertising strategies.

[0165] "Real-time display" refers to the process by which a terminal device, after receiving advertising display data, completes the parsing and rendering process within a predetermined time threshold without significant human intervention, so that the advertising content is presented instantly on the user's display device.

[0166] "Resend" refers to the action of the server regenerating or updating the advertisement display data and transmitting it to the terminal device through the communication link after changes in user behavior information or system updates of asset text, in order to keep the advertisement content on the terminal device synchronized with the latest generated results.

[0167] In one embodiment of the present invention, a server serves as the core computing device running the program of the present invention. The server includes hardware resources such as a multi-core central processing unit, an optional graphics processing unit, main memory, non-volatile storage devices, and a network interface. The server runs the application program on an operating system (e.g., a general-purpose server operating system). The application program uses a general-purpose interpreted programming language (e.g., Python) to implement functions such as data preprocessing, feature extraction, prompt generation, generative artificial intelligence model invocation, and asset text post-processing. When needed, the server invokes a deep learning framework (e.g., a tensor computation framework) to load the generative artificial intelligence model and utilizes the graphics processing unit to perform high-complexity operations such as matrix multiplication and attention calculation.

[0168] In one embodiment of the invention, the server uses a relational database management system to store user-related attribute information and product-related attribute information. The server constructs structured data structures in main memory, such as two-dimensional arrays, key-value maps, and vector lists, to represent user feature vectors, product feature vectors, and relevance scores. The server also uses a caching system to store frequently accessed data to reduce disk access latency, thereby improving data read speed.

[0169] In one embodiment of the invention, the server performs specific data processing on user-related attribute information. The server reads fields such as user age, gender, geographic region, interest tags, browsing history, search history, and purchase history from the database, and constructs a table structure in memory using a data processing library. The server performs standardization operations on numerical fields, for example, subtracting the mean from each value and dividing by the standard deviation to obtain standardized features with zero mean and unit variance. The server uses one-hot encoding or embedding encoding on categorical fields to map categories to high-dimensional sparse vectors or low-dimensional dense vectors. The server performs sliding window statistics on time-series behavioral data, for example, counting the number of views and clicks for different product categories over the past few days, and stores the statistical results as a fixed-length interest intensity vector.

[0170] In one embodiment of the invention, the server performs similar data processing on product-related attribute information. The server reads fields such as product category, price, price range, promotional information, target audience tags, functional tags, and text description from the product database. The server performs a normalization operation on the price field, mapping the price to a floating-point value within a predefined range. The server uses an embedding matrix to map discrete categories and target audience tags into low-dimensional vectors to reduce dimensionality and preserve semantic similarity. The server can generate description vectors for the product text description using a bag-of-words model or a pre-trained text encoding model for subsequent relevance calculations.

[0171] In one embodiment of the present invention, the server calculates the relevance using the aforementioned user feature vector and product feature vector. The server may employ dot product, cosine similarity, or a small feedforward neural network as the relevance calculation module. The server concatenates the user feature vector and product feature vector into a set of input vectors, performs a nonlinear transformation through a multi-layer fully connected network, and outputs a scalar as a relevance score using an activation function at the output layer. The server sorts multiple product candidates based on the obtained relevance scores and selects one or more product candidates with the highest scores as target product candidates for generating the prompt statement. By performing this feature-level relevance calculation on the server side, the server can quickly filter out the product that best matches the current user from a large number of products, thereby reducing the number of calls and computational load required for subsequent generative artificial intelligence model inference, and improving computational efficiency.

[0172] In one embodiment of the present invention, the server generates a prompt statement based on preprocessed user attribute information and product attribute information. The server first reverse-maps key fields in the user feature vector into readable natural language descriptions. For example, it converts "age=30", "gender=male", "interest tags include running and fitness", and "recent search keyword is running shoes" into Chinese text: "30-year-old male, enjoys running and fitness, and has recently searched for running shoes frequently." Simultaneously, the server converts key information in the product feature vector into natural language descriptions, such as: "New professional running shoes, priced at 800 yuan, currently on a limited-time 20% discount promotion, suitable for long-distance running and daily training."

[0173] The server maintains multiple prompt statement templates in memory. Each template is stored as a string and includes variable placeholders. The server fills the templates with the user's natural language description and the product's natural language description, and also includes descriptions of the input and output conditions for the generative AI model, as well as word count constraints. An example of the prompt statements generated by the server is shown below: User information: 30-year-old male, who enjoys running and fitness, and has recently been frequently searching for running shoes and browsing a variety of mid- to high-priced running shoes.

[0174] Product Information: New professional running shoes, priced at 800 yuan, are currently on a limited-time 20% off promotion, suitable for long-distance running and daily training.

[0175] Task: Based on the information above, generate a responsive ad text (asset text) suitable for display on smartphones for this user. Keep the text to 20-30 Chinese characters, highlighting the offer and the running scenario, and include a clear call to action. In this way, the server doesn't simply concatenate several fields into a prompt statement. Instead, it utilizes feature extraction and template control mechanisms to explicitly define the model's task boundaries, output style, and length limitations within the prompt statement, thereby providing controllable constraints on the behavior of the generative AI model at the computer's internal level. Compared to manually writing unstructured prompt statements, this structured prompt statement generation method significantly improves the consistency and predictability of the model's output, reduces the generation of irrelevant or illegal text, and thus achieves a technical improvement in output quality and system stability.

[0176] In one embodiment of the present invention, the server uses a deep learning framework to load a generative artificial intelligence model. The server can use a sequence-to-sequence architecture based on a self-attention mechanism. The model includes multiple encoding layers and multiple decoding layers, each layer containing a multi-head self-attention sublayer and a feedforward neural network sublayer. During the inference phase, the server inputs the prompt statement into the word segmenter, decomposes the text into sub-word units, and maps them to embedding vectors. The server feeds the embedding vectors into the encoder, performs layer-by-layer matrix multiplication, addition, and non-linear activation, and uses multi-head attention to calculate the dependencies between different positions. The server generates asset text step by step at the decoder, controlling the generation process through decoding strategies such as greedy search, top-k sampling, or top-p sampling. At each generation step, the server calculates the probability distribution of the next output token based on the current decoding state, encoder output, and attention weights, and selects the token with the higher probability as the output.

[0177] During the model training phase, the server uses a large-scale text corpus and known high-quality advertising text as training data. The server defines a cross-entropy loss function to quantify the difference between the model's output sequence and the target sequence as a loss value. During training, the server calculates the gradient of each layer's weights using the backpropagation algorithm and updates the weights using an adaptive learning rate optimization algorithm. The server can employ data augmentation strategies, such as synonym replacement, word order transformation, and advertising scene perturbation, to expand the training samples, thereby improving the model's robustness to diverse input prompts. Through these specific settings of the neural network architecture and training process, the server can generate asset text that conforms to the prompt constraints with high accuracy and speed during the generation phase.

[0178] In one embodiment of the present invention, the server performs multi-level post-processing after generating the asset text. First, the server performs prohibited expression detection, matching the generated asset text against a dictionary of sensitive words and prohibited phrases, and using a simple text classification model to identify potentially risky phrases. The server replaces or regenerates any parts detected as violating regulations, thereby ensuring the technical reliability of the output text in terms of compliance. The server then performs information consistency verification, comparing information such as prices, discounts, and product names in the asset text with the current product attribute data in the system. If inconsistencies are found, the text is corrected to the actual values ​​in the database. Finally, the server performs length adjustment, truncating or shortening text exceeding a predefined character limit, and triggering regeneration or supplementary explanations for text that is clearly too short or lacks sufficient information. Through this post-processing chain, the server establishes an automatic error correction and standardization mechanism at the text level, thereby significantly reducing the error rate caused by the randomness of the generation model from a computer technology perspective.

[0179] In one embodiment of this invention, the server embeds the filtered and modified asset text into the advertisement display data structure. The server constructs a data object in memory containing asset text identifiers, product identifiers, material file identifiers, jump links, and display location information, and encodes it as a response message in a serialized format. The server sends this response message to the terminal via a network interface using a communication protocol. Because the number of advertisements to be sent has been reduced on the server side through product candidate filtering and relevance calculation, and the asset text has been compressed and its length controlled, the server can reduce network transmission load while maintaining personalization, thereby reducing communication bandwidth consumption and shortening terminal-side rendering latency, demonstrating the technical effects of communication load reduction and response speed improvement.

[0180] In one embodiment of the present invention, a terminal serves as a display and interaction device, receiving and processing advertising display data. The terminal includes a processor, a memory, a display device, and a network module, and runs an operating system and applications. After receiving the advertising display data from the server, the terminal parses the message structure in its local memory, extracting asset text, image addresses, redirect links, and display layout parameters. The terminal populates the asset text into interface controls, obtains image resources through the network module and caches them locally, and then renders the advertising component on the display device. During rendering, the terminal can adjust the layout according to screen resolution, orientation, and system theme to ensure the readability of the advertisement on different devices. The terminal also binds click, swipe, long-press, and other interactive event listening logic to the advertising component. When the user interacts, the terminal records the relevant events and packages the behavioral information into structured data, sending it back to the server.

[0181] In one embodiment of the present invention, a user, as the user of the system, browses and interacts with advertisements on a terminal. The user can scroll through pages, view advertisement cards, click on advertisements to jump to product details pages, or close the application. These user behaviors are captured through terminal interaction events, forming user behavior information. Because the terminal meticulously records elements such as behavior timestamps, dwell time, and interaction type, the server, upon receiving this information, can make a more granular estimate of the user's interest level. Compared to the traditional simple method of only counting clicks, the behavior collection method of the present invention can provide richer feature inputs for subsequent algorithms, thereby improving the accuracy of interest modeling.

[0182] In one embodiment of the invention, the server processes user behavior information returned by the terminal. The server writes the behavior information to a log system or a real-time processing queue and updates the user profile through an online or near-real-time processing module. The server can use an incremental update algorithm to dynamically adjust the weight distribution of the interest vector based on new click, browsing, and closing behaviors. For example, the server can use an exponential decay function to assign different weights to behaviors in different time periods, making recent behaviors have a greater impact on interest estimation. The server can also construct a short-term interest representation based on the behavior sequence, merge it with the long-term profile, and generate a new user feature vector. By directly applying behavioral feedback to the feature layer and the prompt generation logic, the server implements a closed-loop mechanism for adaptive update strategies within the same technical system, significantly improving the system's response speed and personalization accuracy in the face of changes in user interests.

[0183] In one embodiment of the present invention, the overall module division of the server includes a data acquisition and preprocessing module, a feature extraction and relevance calculation module, a prompt statement generation module, a generative artificial intelligence model inference module, an asset text post-processing module, an advertisement packaging and sending module, and a behavior feedback processing and profile update module. The modules communicate with each other through clearly defined data structure interfaces. The server constructs a directed data flow in memory, allowing data to be transformed sequentially through multiple stages, starting from raw logs and database tables, including feature vectors, candidate sets, prompt statements, generated text, and displayed data. This modular data flow design allows the server to be independently optimized at each stage. For example, vectorization can be introduced to accelerate the feature extraction stage, parallel computing can be utilized in the model inference stage, and efficient string search algorithms can be applied in the post-processing stage, thereby improving the overall performance and scalability of the system in complex data processing and model calling scenarios.

[0184] In other embodiments of the present invention, the server can employ different generative artificial intelligence model structures. For example, the server can use a text generation model based on recurrent neural networks, or a variant model based on an encoder-decoder structure, or a multi-task learning architecture to jointly train ad title generation and ad description generation. The server can also break down the prompt statements into multiple sub-prompts input to different sub-models, thereby generating more granular asset text components. During the feature calculation stage, the server can also replace different relevance models, such as using collaborative filtering models or graph neural networks, to uncover implicit relationships between users and products. These variants, without changing the overall system flow of the present invention, can all achieve similar technical effects, namely, efficient, accurate, and controllable generation of ad asset text through the combination of structured prompt statements and deep learning models within the server.

[0185] Through the above embodiments, it can be seen that the system of the present invention does not simply automate the manual text generation process, but rather constructs a technical framework within the server centered around prompt statements and a generative artificial intelligence model. Through feature-level data processing, structured prompt statement generation, multi-layered post-processing, and a behavior-driven online update mechanism, the server fundamentally improves the computer's capabilities in high-dimensional data management, text generation control, and real-time feedback utilization. This results in increased generation accuracy, faster processing speed, lower error rates, and reduced communication resources, thus achieving substantial improvements at the computer technology level.

[0186] use Figure 12 The processing flow is explained.

[0187] Step 1: The server reads user-related attribute information and user behavior logs from the storage device, and product-related attribute information from the product database. Inputs include basic user information (age, gender, region, interest tags, etc.), user behavior records (browsing, searching, clicking, purchase logs, etc.), and product attribute information (category, price, promotional information, target audience, text description, etc.). The server performs missing value imputation, outlier removal, and format standardization on the raw data, writing the results into a structured data table in memory. Outputs are the cleaned original user data table and the original product data table.

[0188] Step 2: The server performs feature processing on the raw user data table obtained in step 1. The input is the raw user data table. The server normalizes numerical fields such as age, number of purchases, and spending amount using numerical standardization; vectorizes category fields such as gender, region, and interest tags using one-hot encoding or embedding encoding; and calculates the number of views and clicks for different product categories using a sliding time window statistical algorithm to form interest intensity vectors. The server combines these processing results into fixed-length user feature vectors based on user identifiers. The output is a set of user feature vectors generated for each user.

[0189] Step 3: The server performs feature processing on the raw product data table obtained in step 1. The input is the raw product data table. The server performs linear normalization on the price field, mapping the price to a floating-point value within a range; it uses embedding matrices to generate dense vectors for product category and target audience tags; and it calls the text encoding module to convert the product text description into a semantic vector. The server concatenates the price features, category embeddings, tag embeddings, and description vectors according to the product identifier to form a product feature vector. The output is a set of product feature vectors generated for each product.

[0190] Step 4: The server calculates the relevance between the user feature vector from step 2 and the product feature vector from step 3, and then filters candidate products. The input consists of the target user's user feature vector and a set of all or part of the product feature vectors. The server calculates a relevance score between each product feature vector and the user feature vector using dot product, cosine similarity, or a small feedforward neural network, and stores the results in a score list. The server sorts the score list and selects the top-scoring products as a candidate product set. The output is a list of candidate products most relevant to the target user and their corresponding feature vectors.

[0191] Step 5: The server generates natural language descriptions of user profiles and product profiles based on user features and candidate product features. The input includes the target user feature vector, basic user fields, candidate product feature vectors, and basic product attributes. The server maps numerical feature ranges to semantic labels (e.g., age group, consumption level) using rule mapping and converts categories with higher weights in the interest intensity vector into interest description text. The server templates the category, price, promotional information, and target audience for each candidate product, forming the product profile text. The server assembles the user profile text and product profile text into multiple natural language sentences. The output consists of a set of user profile text and the corresponding set of product profile text.

[0192] Step 6: The server constructs prompts for the generative AI model based on the user profile text and product profile text generated in step 5. The input consists of user profile text, one or more product profile texts, and a predefined prompt template. The server selects a suitable prompt template from memory, fills the placeholder positions in the template with the user profile and product profile, and includes explicit descriptions of input conditions, output conditions, and word limits. The server concatenates these parts into a complete natural language text. The output consists of one or more prompts, each explicitly describing user information, product information, and the requirements for the asset text generation task.

[0193] Step 7: The server converts the prompt text generated in step 6 into a model-readable token sequence and inputs it into the generative AI model. The input is the prompt text. The server calls a tokenizer to segment the prompt text into subwords or tokens, and uses an embedding matrix to convert the tokens into a vector sequence. The server feeds this vector sequence into the encoding part of the generative AI model, performing matrix multiplication, addition, and activation function operations in multiple self-attention layers and feedforward network layers to obtain the encoded representation. The output is a context vector representation for decoding.

[0194] Step 8: During the decoding phase, the server invokes a generative AI model to generate asset text. The input consists of the context vector representation obtained in step 7 and the decoding start marker. At each decoding time step, the server calculates the probability distribution of the next marker, selects a specific marker using strategies such as greedy search, top-k sampling, or top-p sampling, appends it to the current output sequence, and inputs the new sequence back into the decoder. The server repeats this process until a termination marker is generated or the preset maximum length is reached. The output is the initially generated asset text sequence, corresponding to natural language advertising copy.

[0195] Step 9: The server performs prohibited expression detection on the asset text generated in step 8. The input is the asset text string. The server uses a string matching algorithm and a sensitive word dictionary to detect whether the asset text contains prohibited words, insults, illegal content, etc. If a match is detected, the corresponding location is recorded. The server can call a simple discriminant model to score the overall risk of the text. The server performs replacement, trimming, or regeneration operations on the parts judged to be in violation. The output is the asset text with the in violation content removed.

[0196] Step 10: The server performs an information consistency check on the asset text obtained in step 9. The input consists of the asset text and the actual attribute information of the current product. The server uses pattern matching and numerical parsing algorithms to extract structured fields such as price, discount, and product name from the asset text and compares them with the corresponding fields in the product database. When discrepancies are found, the server modifies the relevant descriptions in the asset text based on the actual data in the database, reorganizing the sentence structure if necessary. The output is the asset text that maintains consistency with the actual product attributes in both numerical and textual terms.

[0197] Step 11: The server adjusts the length of the validated asset text. The input consists of the asset text and pre-defined upper and lower character limits. The server calculates the text length; if it exceeds the upper limit, it truncates unimportant parts based on sentence boundaries or priority strategies. If the length is significantly below the lower limit and the information is insufficient, the server can use a short template to add supplementary information or re-trigger the generation process. The output is the asset text that meets the character limit and is complete with all information.

[0198] Step 12: The server encapsulates the final asset text into ad display data and sends it to the terminal. Input includes the adjusted asset text, product identifier, creative resource identifier, redirect links, and display location information. The server constructs the ad data object in memory and serializes it into a response message format. The server sends this response to the terminal requesting the ad via a communication protocol through the network interface. Output is the ad display data packet transmitted to the terminal over the network.

[0199] Step 13: The terminal receives and parses the advertising display data sent by the server. The input is the advertising data packet received from the network interface. The terminal calls the parsing module to restore the internal data structure, extracting asset text, image addresses, redirect links, and display layout parameters. The terminal populates the interface text controls with the asset text, loads image resources into image controls via network requests, and arranges the advertisement's position on the screen according to the layout parameters. The output is the advertising interface rendered on the terminal display device.

[0200] Step 14: Users browse and interact with advertisements on the terminal. The input is the advertisement interface displayed on the terminal screen. Users can perform actions such as scrolling, clicking on advertisement cards, clicking the close button, or long-pressing. The specific behavioral events generated (such as impressions, clicks, closes, and dwell times) are recorded by the terminal system. The output is multiple user behavior records with timestamps and event types.

[0201] Step 15: The terminal packages the user behavior records and sends them back to the server. The input is the list of user behavior records generated in step 14. The terminal organizes the behavior records into structured data according to a preset reporting strategy (e.g., real-time reporting or batch reporting) and sends it to the behavior reporting interface specified by the server via the network interface. After receiving confirmation from the server, the terminal marks this reporting as complete. The output is a user behavior data packet successfully transmitted to the server.

[0202] Step 16: The server processes user behavior data returned from the terminal and updates the user profile. The input consists of user behavior data packets and the currently stored user feature vector. The server parses each behavior record, extracting fields such as event type, event time, associated products, and dwell time, and writes them to the behavior log storage. The server uses an incremental update algorithm to weight and adjust the interest intensity vector, assigning higher weight to recent behaviors and attenuating older behaviors. The server calculates the updated user feature vector based on the new behavioral characteristics and writes this vector back to the user feature storage module. The output consists of the updated user feature vector and the user profile, providing input for the next round of prompt generation and asset text generation.

[0203] Alternatively, an emotion engine for inferring user emotions can be combined. That is, the specific processing unit 290 can also use the emotion-specific model 59 to infer user emotions and perform specific processing using user emotions.

[0204] Example 2 The flow of a specific process in Example 2 will be described. Each part of the system described below is implemented by the data processing device 12 and the smart device 14. The data processing device 12 will be referred to as the "server," and the smart device 14 as the "terminal."

[0205] In existing ad generation and delivery technologies, ad text is typically pre-written manually or generated using simple template rules, which presents the following technical problems: First, when processing ad-related data, servers can only combine product information and user attributes based on static rules, lacking the ability to understand and express complex semantics, making it difficult to finely customize ad text for different users; second, although servers can collect user behavior data, they lack the technical mechanism to dynamically couple this behavior data with the text generation process, failing to fully utilize multi-dimensional information such as user history and real-time feedback when generating ad copy, often requiring manual intervention for ad optimization, resulting in low overall system automation and response speed; third, existing systems... In current systems, calling generative AI models typically involves simply passing user input directly to the model. This lacks a server-side computational process for structuring and integrating the input information, extracting features, dynamically constructing prompts, and iteratively updating them. This results in the model output failing to consistently meet the multiple constraints of advertising operations (such as placement scenarios, target audiences, style limitations, and length restrictions), thus limiting the technical effectiveness of generative AI models in advertising generation scenarios. Furthermore, after receiving user feedback, the server usually only modifies the recorded advertising content, without incorporating the feedback information as internal input into the prompt construction and generative AI model inference. This lack of a closed-loop learning and adaptive optimization architecture is problematic. Therefore, a technical solution is needed that can uniformly model information characteristics, user attributes, user behavior history, and real-time feedback on the server side. This solution should dynamically construct prompts, call generative AI models to generate advertising text, automatically update advertising information, and learn and improve based on user feedback and behavioral data. This would improve the computer's processing power and efficiency in the automatic generation and optimization of advertising text from both computational and system architecture perspectives.

[0206] The specific processing performed by the specific processing unit 290 of the data processing apparatus 12 in Embodiment 2 is achieved by the following means.

[0207] In this invention, the server includes means for acquiring input information containing information characteristics and user attributes, and further acquiring conditional information related to advertising placement; means for performing feature extraction and combination on the input information, conditional information, accumulated user behavior history, and advertising response information in the processing unit within the server, and constructing prompt statements as input to a generative artificial intelligence model; means for inputting the prompt statements into the generative artificial intelligence model to generate advertising text information and processing its format; means for sending the generated advertising text information to an advertising management device and triggering automatic updates of the content area in the advertising information; means for acquiring updated advertising information from the advertising management device and outputting it to a user terminal through a display device; and means for receiving user confirmation results or correction instructions for advertising content through the user terminal, storing the results or instructions as acceptance information along with user behavior history as accumulated information, generating feature quantities representing information characteristics and user attributes based on the accumulated information, and appending the feature quantities to subsequent prompt statements to repeatedly execute the generation and update process. This allows for the formation of a closed-loop data processing chain on the server side, centered on the dynamic construction of prompt statements. This tightly integrates the invocation of generative artificial intelligence models with user attributes, behavioral history, real-time feedback, and advertising delivery conditions, thereby enhancing the computer system's processing capabilities in automatic advertising text generation, personalization, and real-time optimization. It also reduces reliance on manual editing and rule configuration, improves the automation and response speed of advertising generation and updates, and ultimately enhances the overall utilization efficiency of computing resources and the user-side display effect.

[0208] A "system" refers to an overall device and operating environment consisting of multiple interconnected hardware and / or software modules, used to perform a series of functions such as information acquisition, data processing, model reasoning, result output, and feedback processing.

[0209] "Information processing device" refers to a device with data input, storage, calculation, control and output functions, including but not limited to servers, computer terminals or clusters thereof, used to carry various programs and perform corresponding processing in this invention.

[0210] The "processing unit" refers to a logical functional unit located in an information processing device, consisting of a processor and its controlled software program, used to perform computational processing such as input information parsing, feature extraction, prompt statement construction, model calling, result processing, and feedback learning.

[0211] "Information characteristics" refer to the data attributes and descriptive information associated with the advertising target, display scenario, and placement conditions, including the category, function, selling points, activity type, placement channel, and placement information of the goods or services, which are used to characterize the features of the content to be promoted and its environment.

[0212] "User attributes" refer to static or relatively stable characteristic information related to users, including but not limited to age group, gender, geographical region, interests, occupation type, consumption level, etc., which are used to characterize the target user group in the process of ad generation and optimization.

[0213] "Input information" refers to the data set acquired by the processing department and used for subsequent calculations and processing, including information characteristics, user attributes, advertising placement conditions, input from user terminals, and historical behavior data.

[0214] "Ad placement conditions" refers to the constraints and configuration information related to the ad display strategy, including the placement period, placement channels, display positions, budget limits, target conversion type, copy length or style requirements, etc.

[0215] "Generative artificial intelligence models" refer to artificial intelligence models built on machine learning and deep learning technologies that can automatically generate text information based on input data, including but not limited to language models, dialogue models, or text generation models based on neural networks.

[0216] "Prompt statements" refer to indicative texts constructed by the processing department and used as input to generative artificial intelligence models. They are used to explicitly provide the model with generation goals, content scope, style requirements, and constraints, thereby guiding the model to generate advertising text information that meets expectations.

[0217] "Advertising text information" refers to one or more pieces of text information automatically generated by generative artificial intelligence models based on prompts, used to constitute advertising content, including text content such as titles, body text, descriptions, and promotional slogans.

[0218] "Advertising management device" refers to a device or system used to manage advertising-related data and business processes, including functional modules such as the storage, editing, version control, delivery configuration and update control of advertising materials, used to receive and apply text information for advertising to form displayable advertising information.

[0219] "Advertising information" refers to the overall collection of advertising data that can be displayed and placed, including text information, images, audio and video, links, display styles, and placement parameters.

[0220] "Content area" refers to the logical field or display location in an advertisement that is used to present the text information of the advertisement, including the advertisement title area, body area, description text area, etc., which are used to display text content on the user terminal interface.

[0221] "Display device" refers to a display component or terminal device used to output and present advertising information, including displays, monitors, mobile terminal screens, etc., which can display updated advertising information to users in a visual form.

[0222] "User terminal" refers to terminal equipment used by users to operate and view advertising information, including mobile terminals, tablet devices, desktop computing devices, or other devices with network communication and display functions.

[0223] "Confirmation result" refers to feedback information from users after viewing the advertisement content, indicating whether the content meets their expectations. This includes statuses such as "approved," "rejected," and "needs modification," and is used to guide subsequent generation and optimization processes.

[0224] "Modification instructions" refer to the user's instructions regarding modifications or preferences for advertising content, including style adjustments, content additions or deletions, tone changes, length alterations, etc., which are used as one of the input conditions for constructing subsequent prompt statements and generating text.

[0225] "Acceptance information" refers to the general term for feedback information related to advertising content received and recorded by the processing department from user terminals, including confirmation results, correction instructions, and other evaluative or directive information.

[0226] "Accumulated information" refers to the collection of historical data related to user behavior, ad display and feedback that is continuously collected and stored during system operation. This includes user behavior history, ad response information, acceptance information, etc., which are used for subsequent feature generation and learning-based improvement.

[0227] "User behavior history" refers to the behavioral records generated when users interact with advertisements or related content within a certain time range, including the number of views, clicks, dwell time, conversions, and closing actions.

[0228] "Ad response information" refers to data that reflects users' response to the effectiveness of ad display, including impressions, click-through rate, conversion rate, interaction frequency, and statistical indicators calculated from them.

[0229] "Features" refer to numerical or structured data, including vectors, scalars, and label sets, which are calculated by the processing department based on accumulated and input information and used to quantitatively represent information characteristics and user attributes. They are used as part of the input to generative artificial intelligence models.

[0230] "Status information" refers to a set of data that can represent the current status of users, advertisements, and system environment, including the user's latest input information, user attributes, behavioral history, current session context, and advertisement delivery environment.

[0231] "Real-time execution" refers to the processing department's immediate update of prompt statements, model calls, and advertising information upon detecting changes in status information or within a predetermined short time period, to ensure that advertising content can be generated and displayed in near real-time response time.

[0232] In this invention, the server operates as the core information processing device, and the terminal operates as a user interaction and display device, through which the user inputs and provides feedback information. The following detailed description of the invention's implementation, considering its hardware, software components, and data structures, is provided.

[0233] I. Overall Hardware and Software Composition of the System The server runs on one or more general-purpose computing devices, which may include servers with multi-core central processing units, compute nodes with graphics processing units, and network storage devices. The server may use x86-based server hardware and be equipped with graphics processing units, such as graphics accelerator cards with multiple parallel computing cores and large amounts of video memory. The server may run a general-purpose operating system, such as a Linux kernel-based server operating system.

[0234] At the software level, a server can run web server components (such as reverse proxy and application server software based on HTTP / HTTPS protocols), application runtime environments (such as scripting language runtime environments or virtual machine runtime environments), deep learning inference frameworks (such as frameworks supporting tensor operations and neural network inference), database management systems (such as relational database management software), and logging and monitoring components. The server loads generative artificial intelligence models into memory and performs inference computations on these models through the tensor operation interfaces provided by the deep learning framework.

[0235] The terminal can be a smartphone, tablet, or personal computing device, running a mobile or desktop operating system. The terminal can communicate with the server via HTTPS through a browser application or native applications to enable prompt input, advertisement display, and feedback submission. The browser or application running internally on the terminal uses the network stack and graphical interface library provided by the operating system to perform interface rendering and user input collection locally.

[0236] Users input natural language text through the terminal's input interface, read the advertising text returned by the server, and input corrections or new prompts when needed.

[0237] II. Generative Artificial Intelligence Model Structure and Training Methods The generative artificial intelligence model loaded by the server in this invention is one implementation of generative artificial intelligence models, and can adopt an autoregressive language model based on the Transformer architecture. This model typically includes a word embedding layer, a multi-layer Transformer encoder-decoder structure or a pure decoder structure, and an output projection layer.

[0238] During model initialization, the server maps the vocabulary to a dense vector space and converts the input labeled sequence into a floating-point tensor using a word embedding matrix. The server then adds the positional encoding to the word embedding vector to form an input representation containing positional information. In each Transformer layer, the server performs multi-head self-attention operations and feedforward neural network operations, specifically including: The server computes the query, key, and value matrix, and performs a weighted summation of the similarities between tags at different locations. The server stabilizes gradient propagation between layers through normalization and residual connections. The server performs non-linear transformations in the feedforward network using non-linear activation functions (such as modified linear units or similar functions). At the output layer, the server maps the hidden states to the vocabulary dimension through a linear transformation and uses the softmax function to obtain the probability distribution of each possible next tag.

[0239] The server can use large-scale natural language corpora as training data during the model training phase. It trains the model using maximum likelihood estimation, employs cross-entropy loss as the error function, and uses the difference between the true one-hot encoding of the next label and the predicted distribution as the training error. During training, the server uses stochastic gradient descent algorithms (such as adaptive moment estimation optimization algorithms) for weight updates and calculates the gradient values ​​of each layer of the neural network based on backpropagation. The server can improve the model's generalization ability through data augmentation strategies (such as reorganizing sentence structures in training samples and replacing synonyms).

[0240] In this embodiment of the invention, the server can employ a basic language model pre-trained on a general corpus, and then fine-tune it using advertising text data. During the fine-tuning phase, the server introduces specialized corpus containing product selling point descriptions, promotional copy, user feedback and evaluations, enabling the model to possess better semantic expression and style control capabilities in the advertising domain.

[0241] III. Data Structures and Feature Representation The server maintains various data structures in memory and database to support the construction of prompt statements, model invocation, and advertising information updates. The server can abstract data related to the advertising object into "information characteristic" records, user-related data into "user attribute" records, and user behavior into "user behavior history" records.

[0242] The server can store information characteristics using, for example, the following logical field sets: product category field, function tag field, activity type field, distribution channel field, display position field, and restriction condition field. The server can encode these fields as discrete or sparse features and embed them as needed for constructing prompts or as additional explanations for model input.

[0243] The server can represent user attributes as records including fields such as age group, gender, interest tags, and historical spending levels. The server can perform multi-hot encoding on interest tags and map discrete variables such as age group to a vector space through an embedding matrix, thereby obtaining numerical feature quantities.

[0244] The server can represent user behavior history as sequential data containing fields such as timestamps, ad identifiers, exposure markers, click markers, dwell time, and conversion events. The server can summarize user behavior characteristics over a recent period using a sliding time window, such as click-through rate, average dwell time, and frequency of interaction with different product categories, and use these statistical results as part of the feature set.

[0245] In the feature generation module, the server combines information characteristics, user attributes, and user behavior history to construct high-level features that control the output style and content of the generative artificial intelligence model. In one implementation, the server can embed these features into prompt statements in natural language, for example, by expressing them in the form of sentences describing "target users' preferences for genuine word-of-mouth and practical functions." In another implementation, the server can encode these feature values ​​as additional tags and annotate them in the prompt statements using specific markers.

[0246] IV. Construction and Processing of Prompt Statements After receiving the initial prompts and related input information from the terminal, the server parses and reassembles this input through a program module. The server runs a piece of logic in the application layer program to comprehensively express information characteristics, user attributes, advertising conditions, and the user's original prompts at the natural language level.

[0247] For example, the server can display a user's original prompt statement: "Please generate an advertising copy that highlights the performance advantages of the new product, in a natural and professional tone, and keep it to around 50 words." Combined with product information and campaign context information retrieved from the database by the server, the following prompt statement is generated: "You are an advertising copywriting expert."

[0248] Please generate a Chinese ad copy for display as a banner on the mobile homepage based on the following criteria: - Product Type: Smart Home Devices - Core features: Automatic route planning, remote control, and silent operation. - Target users: Urban office workers aged 25-34 who are interested in smart homes - Copywriting requirements: Natural and professional tone, highlighting time-saving, labor-saving, and intelligent experience, with a word count of around 50 words.

[0249] User prompt: 'Please generate a short advertising copy that highlights the performance advantages of the new product, using a natural and professional tone, and keeping it around 50 words.' The output should only contain the main body of the advertising copy; do not include any descriptions. During the above processing, the server performs specific string concatenation, character encoding conversion, length limit checks, and sensitive content filtering on the input data. The server uses regular expressions and dictionary rules to filter and prune user input to prevent illegal or inappropriate content from entering the inference calculations of the generative artificial intelligence model, thereby improving the overall robustness and output quality of the system.

[0250] In one variant implementation, the server can construct different template-style prompts for different ad placements. For example, when the terminal informs that the current ad placement is the application launch page, the server can add a requirement for "short-term strong appeal" to the prompt; when the terminal informs that it is a long article details page, the server can add a requirement for "detailed explanation of functional features and usage scenarios" to the prompt.

[0251] V. Generation and Post-processing of Advertising Text Information After constructing the prompt statement, the server inputs it as a text sequence into the generative artificial intelligence model. The server converts the prompt statement into a labeled sequence through the inference interface of the deep learning framework and performs forward computation on the graphics processing unit. During the inference phase, the server can set hyperparameters such as maximum generation length, temperature parameters, and top-k or top-p sampling parameters to control the diversity and stability of the output text.

[0252] The server obtains a piece of advertising text information during the model output phase. For example, after the above prompt is input into the model, the server can obtain the following output: "This smart home device automatically plans the route throughout the house, can be remotely controlled with one click, and operates quietly and efficiently, allowing you to easily keep your home clean in your busy life." After receiving the output, the server performs post-processing, including removing leading and trailing whitespace, removing unnecessary prefixes such as "The text is as follows:", limiting the maximum number of characters, and standardizing punctuation. If necessary, the server can trim or regenerate the text based on the space limitations of the ad slot, and store the final ad text information as a structured field in the database of the ad management device.

[0253] In an alternative implementation, the server can control the diversity of model output by generating multiple candidate ad texts at once and then filtering them using a simple scoring function (e.g., based on content length, keyword coverage, and prohibited word detection results). The server uses the candidate with the highest score as the final ad text, thereby improving text quality and compliance while ensuring generation speed.

[0254] VI. Updating of Advertising Management Devices and Control of Terminal Display After generating the text information for the advertisement, the server sends the text along with the advertisement identifier to the advertisement management device. The server performs a write operation through the database access module, writing the text into the corresponding advertisement content field. Simultaneously, the server records the version number, generation time, and the identifier of the generative AI model used, for subsequent version backtracking and performance analysis.

[0255] After receiving the advertising information from the server, the terminal uses a graphical user interface library to render it locally. The terminal fills the advertisement with text information into preset text controls, loads image links as bitmap resources, and displays them on the screen in a specified layout. The terminal can also adjust the display method according to style parameters provided by the server (such as font size, line count limits, etc.) to ensure readability on different device resolutions.

[0256] After viewing the advertisement content on the terminal, users can perform actions using the "Confirm" or "Modify" buttons on the interface. If the user chooses to modify, they can directly enter a new prompt in the terminal's input box, for example: "Please reduce exaggerated descriptions, use a more objective tone, and emphasize the real user experience." The terminal sends the new text along with the advertising identifier to the server. The server then constructs a new prompt and invokes a generative artificial intelligence model to complete a new round of text information generation and updating for advertising.

[0257] VII. Feature Utilization and Learning Improvement The server continuously records user behavior data during operation. Every time a user clicks on an advertisement, navigates to a product details page, stays on the page for a certain period of time, or closes an advertisement, it is recorded as a behavioral event through communication between the terminal and the server. The server then archives and indexes these events in its database according to user and advertisement identifiers.

[0258] In the feature generation module, the server calculates statistical features for each type of ad placement and each target user group, such as the click-through rate of different copy styles over a period of time and the conversion rate of copy containing certain keywords. The server uses these statistical features and processing information (such as users repeatedly requesting "objective tone" and "not overly exaggerated") to construct a preference vector.

[0259] When constructing subsequent prompt statements, the server embeds these preference vectors into the prompt statements in natural language. For example, the server adds the following description when constructing the prompt statements: "Based on historical data, this target user group prefers authentic word-of-mouth style and descriptions of practical functions. Please add expressions that are close to user reviews in your copy." The server transforms features into text constraints that align with the understanding methods of generative artificial intelligence models, thereby improving the matching degree between generated results and user preferences without altering the model's internal structure. Through this improvement based on statistical features and feedback loops, the server essentially optimizes the internal prompting algorithm and feature fusion strategy, thus enhancing the accuracy and stability of automatically generated text.

[0260] VIII. Technical Effects and Improvements in Computer Technology By internally executing specific data structure designs and prompt statement construction processes, the server transforms the invocation of generative artificial intelligence models from simply accepting user natural language input to performing structured integration based on multi-source data. The server achieves fine-grained control over model input by characterizing and combining information features, user attributes, user behavior history, and processed information. The resulting technical effects include: When generating text information for advertisements, the server can more accurately meet length constraints, style constraints, and scene constraints, thereby reducing the number of invalid generation attempts, improving the efficiency of generation results, reducing the consumption of inference resources in the graphics processing unit, and improving the overall processing speed.

[0261] By continuously learning and improving user behavior history and ad response information, the server optimizes the mapping relationship between feature quantities and prompt statements, thereby achieving higher click-through rates and conversion rates under the same hardware conditions. This is reflected in the improved accuracy of the automatic generation system in the internal calculation task of "predicting user preferences".

[0262] The server uses a centralized advertising management system and version control, so changes to advertising copy only require updating a few fields in the database, while the terminal only needs to receive structured data and render it. This design reduces the logical complexity and data transmission volume on the terminal side, lowers the communication load, and improves the utilization efficiency of network resources.

[0263] The server employs a generative artificial intelligence model for semantic-level generation, rather than traditional template replacement or keyword concatenation. It utilizes a self-attention mechanism to model long-distance dependencies during the generation process, effectively reducing semantic errors and expressive redundancy caused by template rigidity. Technically, the server leverages the distributed representation and multi-head self-attention characteristics within neural networks to achieve smoother error distribution and more stable gradient propagation at the computational level, indirectly improving the robustness of the inference process.

[0264] The server constructs and attaches unconventional rules to prompt statements and features, such as creating different preference descriptions for different target metrics (click-through rate, conversion rate), forming a machine-executable rule set that differs from human handwriting. These rules are encoded into the prompt statements in a machine-readable manner, directly influencing the neural network's inference trajectory. This represents a technical control method that utilizes the internal representation space of artificial intelligence, rather than simply automating human writing operations.

[0265] In these implementations, the server optimizes the internal information processing flow of the computer through specialized data structure design, feature generation algorithms, prompt construction strategies, and model inference control. This results in improved accuracy of the advertisement generation task, shorter response time, reduced resource consumption, and enhanced data management efficiency. Therefore, the system described in this invention is not limited to efficiency improvements at the business process level, but provides an improvement solution in the operating mechanism and resource utilization of computer technology itself.

[0266] use Figure 13 The processing flow is explained.

[0267] Step 1: Users enter prompts and submit requests through the terminal.

[0268] Users type natural language text into the input interface of the terminal, such as "Please generate an advertising copy that highlights the performance advantages of the new product, with a natural and professional tone, and keep it to around 50 words.", and then click the "Generate" button.

[0269] The terminal uses the prompt statement and user attributes (such as age group, interest tags, etc.) stored locally on the terminal as input data. It uses the network library provided by the operating system to encapsulate the above data into a request message, performs UTF-8 encoding and JSON encapsulation, and then sends it to the interface address specified by the server via HTTPS.

[0270] During this process, the terminal performs basic validation on the user input (non-empty check, length limit) as a simple data processing step to ensure that the data submitted to the server is in the correct format.

[0271] Step 2: The server receives the request and parses the input information.

[0272] The server takes the HTTP request message sent by the terminal as input, parses the HTTP header and request body through the web server component, extracts the JSON string and converts it into an internal data structure.

[0273] The server parses fields such as prompts, user attributes, and ad placement identifiers from the input data, performs type checks and length validation on each field, and filters out illegal characters; this is the structuring and cleaning process of the input data.

[0274] The server outputs a standardized input information object, including standardized prompt text and structured user attribute data, providing a foundation for subsequent feature extraction and prompt construction.

[0275] Step 3: Server information acquisition characteristics and advertising delivery conditions.

[0276] The server takes the ad placement identifier and product identifier obtained in step 2 as input, and sends a query command to the ad management database through the database access module to obtain information characteristics and ad placement conditions related to the ad, such as product category, core function, activity type, placement channel and display position.

[0277] The server performs join operations and field selections on records from multiple tables retrieved by the query, merging information scattered across different data tables into a single logical record; this process belongs to relational operations and data integration.

[0278] The server outputs an information characteristic and delivery condition object, which includes product feature fields, activity fields, display location fields, etc., to provide context data for constructing prompt statements.

[0279] Step 4: The server generates feature values ​​and integrates user behavior history.

[0280] The server takes user and ad identifiers as input and retrieves user behavior history and ad response information from the behavior log database, including historical exposure counts, click records, dwell time, conversion events, etc.

[0281] The server performs time window aggregation, counting statistics, and ratio calculations on the original behavioral event sequence. For example, it calculates the user's click-through rate and average dwell time for related product categories in the past 30 days; this is a statistical feature extraction operation.

[0282] The server jointly encodes behavioral statistics results with information characteristics and user attributes to generate feature vectors or tag sets representing user preferences and advertising context. The output is a high-dimensional feature object used for subsequent prompt statement construction and model control.

[0283] Step 5: The server constructs prompts for generative artificial intelligence models.

[0284] The server takes the standardized prompt statement from step 2, the information characteristics and delivery conditions from step 3, and the feature quantities generated from step 4 as input, and constructs a structured prompt statement in memory through operations such as string concatenation, template filling, and condition selection.

[0285] The server selects different constraint expressions based on the content of the feature value. For example, when the feature value indicates that the user prefers an objective evaluation, the server adds the explanation "objective tone, increase the description of real user experience" to the prompt statement; this is rule selection processing based on feature value.

[0286] The server outputs a complete prompt text, such as: "You are an advertising copy expert. Please generate a Chinese advertising copy for a mobile homepage banner display based on the following conditions: ... The output is only the advertising copy text, do not include any explanation.", as direct input to the generative artificial intelligence model.

[0287] Step 6: The server performs security and compliance filtering on the prompt statements.

[0288] The server takes the prompt generated in step 5 as input, uses a pre-configured sensitive word dictionary and regular expression rules to perform a matching scan on the text content, and identifies any inappropriate words or illegal expressions that may be contained.

[0289] When the server detects inappropriate content, it performs replacement or deletion operations and truncates excessively long segments if necessary; this is a string matching and replacement operation used to ensure the compliance of model input.

[0290] The server outputs a message indicating that the filter has been passed, and simultaneously logs the filtering process to ensure that subsequent generation processes are conducted within a safe and controllable range.

[0291] Step 7: The server calls a generative artificial intelligence model to generate text.

[0292] The server takes the prompt from step 6 as input, converts the text into a labeled sequence using a deep learning inference framework, maps it to a floating-point vector tensor using an embedding matrix, and performs forward inference computation of the Transformer network on the graphics processing unit.

[0293] During inference, the server performs self-attention calculations and feedforward network operations at each time step, selecting the next marker based on the softmax probability distribution until the set maximum length is reached or an end marker is generated; this is a recursive probability sampling and matrix operation process.

[0294] The server outputs a generated advertising text message, such as "This smart home device automatically plans the route throughout the house, can be remotely controlled with one click, and operates quietly and efficiently, allowing you to easily keep your home tidy in your busy life.", as the result text of the model's inference.

[0295] Step 8: The server performs post-processing on the generated text information of the advertisement.

[0296] The server takes the original advertising text generated in step 7 as input and performs string cleaning, length control and format normalization processing, including removing redundant prefixes and suffixes, merging duplicate punctuation, and limiting the number of characters to a predetermined range.

[0297] The server can truncate or rewrite text based on the maximum number of characters in the ad slot: if the text length exceeds the threshold, the server triggers a regeneration after parameter adjustment, or obtains a more natural short text through intelligent truncation logic (such as truncating at sentence boundaries); this is conditional judgment and text trimming operation.

[0298] The server outputs a text message for the advertisement that meets the display requirements, which is then written as the content field value to the advertisement management device.

[0299] Step 9: The server updates the advertising content in the advertising management device.

[0300] The server takes the advertising text information and advertising identifier from step 8 as input, performs an update operation through the database access module, writes the text into the content area field of the advertising information record, and updates the version number and timestamp.

[0301] The server performs an insert operation on the historical records of the same ad ID, saving the old version of the text and achieving version control; this is a dual operation of database update and insert.

[0302] The server outputs an updated record of advertising information, which includes the latest content text and its metadata, for subsequent distribution to the terminal for display.

[0303] Step 10: The server returns updated advertising information to the terminal and controls its display.

[0304] The server takes the updated advertising record from step 9 as input, encapsulates fields such as title, content text, image link, and redirect link into a response message, and sends it to the corresponding terminal via HTTPS.

[0305] The terminal takes the response message as input, parses the JSON structure data, assigns the text information of the advertisement to the text property of the local UI component, passes the image link to the image loading module and draws it on the screen; this is the data parsing and interface rendering process.

[0306] The terminal outputs a new advertisement displayed on the screen for users to view and interact with.

[0307] Step 11: Users can confirm the advertising content or request corrections through their devices.

[0308] The user takes the advertisement screen shown in step 10 as input, reads the advertisement text on the terminal interface and makes a judgment. If the copy is deemed appropriate, the user clicks the "Confirm / Approve" button. If the copy is deemed inappropriate, the user clicks the "Modify" button and enters a new natural language instruction in the input box, such as "Please reduce exaggerated descriptions, make the tone more objective, and emphasize the real user experience."

[0309] The terminal takes the user's operation result and any new prompt text that may be entered as input data, encapsulates it into a feedback message, and sends it to the server via HTTPS; this is the event collection and feedback packaging process.

[0310] The terminal outputs a feedback request, including confirmation results or correction instructions, to provide a basis for subsequent learning updates to the server.

[0311] Step 12: The server receives user feedback and updates accumulated information and feature values.

[0312] The server takes the feedback message sent by the terminal in step 11 as input, parses out the confirmation result or correction instruction, and writes it together with the current ad ID, user ID and timestamp into the behavior and feedback database.

[0313] The server merges and statistically analyzes new feedback information with existing user behavior history, and updates the user's preference features corresponding to the advertising category, such as adding tags like "prefers objective tone" and "prefers user experience description"; this is based on incremental statistics and tag update processing of time series data.

[0314] The server outputs updated cumulative information and a new set of features, and appends these features to the input when constructing prompts, thereby automatically adjusting the copy style and content focus in the next generation loop.

[0315] Application Example 2 The process flow corresponding to the specific processing in Use Case 2 will be described below. The various parts of the system described below are implemented by the data processing device 12 and the intelligent device 14. In addition, the data processing device 12 is referred to as the "server" and the intelligent device 14 is referred to as the "terminal".

[0316] In existing ad generation and delivery technologies, servers typically generate ad text by combining product information and user attributes based on fixed templates or simple rules. This rule-based generation method has the following problems: First, servers lack the ability to comprehensively model multi-source data when constructing ad text, failing to fully utilize dynamic information such as user behavioral history and emotional state data, resulting in limited ad personalization and poor overall system recommendation performance. Second, server-side ad generation processes are often offline or triggered infrequently; once generated, ad text is used for a long time and cannot be continuously updated based on real-time changes in user behavior and emotions, making it difficult for the system to adjust content in a timely manner during operation and resulting in inefficient use of computing resources. Third, although servers can call certain natural language processing methods... While the system is modular, it lacks a systematic mechanism for constructing and iterating prompts, making it impossible to dynamically optimize prompts based on ad performance data (such as click-through rate and conversion rate). This hinders the full utilization of the expressive power of generative AI models, resulting in low efficiency in the computation process for hardware and network resources. Fourth, multimodal data such as images, audio, and text collected by the terminal are often processed in isolation in existing systems. Servers lack a unified data processing and feature extraction pipeline, preventing the integration of sentiment analysis results with historical behavioral data for online adjustments to ad generation logic. This leads to lengthy processing paths and high latency in high-concurrency scenarios.

[0317] Therefore, a new system architecture and data processing method are needed to enable the server to: uniformly manage and parse product information, user information, behavioral history data, and emotional state data; automatically construct high-quality prompt statements and call generative artificial intelligence models to generate advertising text data based on this; perform closed-loop optimization of prompt statements and advertising text data in combination with advertising performance indicators; and update the advertising composition data in real time during runtime, thereby improving advertising relevance while improving the server-side computing process, data flow, and resource utilization efficiency, and thus improving the overall system performance and scalability from a computer technology perspective.

[0318] The specific processing performed by the specific processing unit 290 of the data processing apparatus 12 in Application Example 2 is achieved by the following means.

[0319] In this invention, the server includes a module for acquiring product information and user information as input data and performing standardized processing; a module for constructing prompt statements based on the product information and user information and inputting them into a generative artificial intelligence model; a module for inputting the prompt statements into the generative artificial intelligence model to automatically generate advertising text data; a module for embedding the generated advertising text data into advertising composition data; a module for outputting the advertising composition data to a display device to present advertising information; a module for collecting and storing user operation information and browsing information as behavioral history data in a data storage unit; and a module for... The system comprises several modules: a module for calculating the performance metrics of the advertising text data and automatically adjusting the composition conditions of the prompt statements based on the calculation results, thereby re-calling the generative AI model to update the advertising text data; a module for acquiring user state data through image and sound acquisition devices and performing sentiment analysis to determine sentiment state data; and a module for generating prompt statements corresponding to the sentiment state based on the determined sentiment state data, along with product information and user information, and inputting them into the generative AI model to automatically generate advertising text data adapted to the sentiment state, and using the advertising text data to update the advertising composition data in real time. This creates a unified processing and generation chain for multi-source data on the server side. The prompt statement-driven generative AI model dynamically generates and iterates advertising text data, enabling real-time personalized updates of advertising content. Simultaneously, closed-loop optimization of the prompt statements and generation results is performed using behavioral history data and sentiment state data, thereby shortening the advertising generation path, reducing server computation latency, improving model call efficiency and data processing throughput, and fundamentally improving the technical performance of the computer system in personalized advertising generation and display scenarios.

[0320] A “system” refers to a whole consisting of one or more information processing devices, storage devices, communication devices, and hardware and software for executing programs, which is a combination of technologies used to realize the generation and display of advertisements.

[0321] "Product information" refers to attribute data related to the promoted object, including product category, name, function, price range, target audience, and other structured or unstructured information used to describe product characteristics.

[0322] "User information" refers to attribute data related to the recipients of advertisements, including age, gender, interests, region, purchase history, and other information used to characterize user features.

[0323] "Generative artificial intelligence model" refers to a text generation model trained based on machine learning and deep learning techniques. It takes prompts as input and outputs natural language text data that meets semantic requirements.

[0324] "Prompt statements" refer to the instructional text input into the generative artificial intelligence model. They are used to explain the generation goals, style, length limits, and context related to product and user information, in order to guide the generative artificial intelligence model to generate text data for the target advertisement.

[0325] "Advertising text data" refers to natural language text automatically generated by generative artificial intelligence models based on prompts, used to constitute advertising content, including headlines, body text, button text, and other textual elements.

[0326] "Advertising composition data" refers to a comprehensive set of data used to present advertising information on display devices, including text data, image data, layout information, interactive element configuration, and related parameters used for display control.

[0327] "Display device" refers to an output device that can receive advertising composition data and present advertising information to users in a visual manner, including displays, monitors, mobile terminal screens and other image display devices.

[0328] "Operation information" refers to the interactive data performed by users on the terminal in response to advertisements or interfaces, including records of operation behaviors such as clicking, swiping, scrolling, inputting, and closing.

[0329] "Browsing information" refers to data related to a user's access to and dwell time on advertising or content pages on a terminal, including page visit time, dwell time, scroll depth, and related session identifiers.

[0330] "Behavioral history data" refers to the collection of operational and browsing information accumulated over a period of time, used to reflect the user's interaction trajectory and behavioral patterns with advertisements and related content.

[0331] "Performance metrics" refer to quantitative parameters used to evaluate the effectiveness of text data or ad composition data in advertising, including impressions, click-through rate, conversion rate, dwell time, interaction rate, and other statistical data used to measure ad performance.

[0332] "Image acquisition device" refers to a hardware device used to acquire images of a user's facial expressions or a scene, including cameras, cameras, and other sensing devices with image acquisition capabilities.

[0333] "Sound acquisition device" refers to a hardware device used to acquire audio signals such as user voice and sound environment, including microphones, voice acquisition modules, and other devices with sound acquisition functions.

[0334] "User status data" refers to the raw or preprocessed data acquired and representing the user's current status through image acquisition devices and sound acquisition devices, including facial images, voice signals, and feature data extracted from them.

[0335] "Sentiment analysis processing" refers to the process of processing user state data or text data to identify the types and intensities of emotions contained therein, usually using pattern recognition, natural language processing or machine learning methods.

[0336] “Emotional state data” refers to structured data obtained through sentiment analysis that represents the category and intensity of a user’s emotions, such as emotion labels like joy, dissatisfaction, and neutrality, and their confidence levels.

[0337] "Emotion-related text data" refers to text data that contains user emotional expressions, including user evaluation information, feedback information, message content, and other natural language text that can reflect user emotions.

[0338] "Evaluation information" refers to textual or symbolic information such as opinions, ratings, and comments given by users regarding products, services, or system functions, used to express their level of satisfaction or dissatisfaction.

[0339] "Response information" refers to the text replies, selection results, or other forms of feedback data that users make after receiving advertisements or system prompts.

[0340] "Real-time updates" refers to a processing method in which, during system operation, when behavioral history data or emotional state data changes, the system automatically regenerates or adjusts the text data for advertising within a short period of time, and immediately reflects this in the advertising composition data and displayed content.

[0341] The embodiments of this invention will be described focusing on the collaborative processing of the server, terminal, and user in terms of hardware structure, software modules, and data flow. The system of this invention does not remain at the level of abstract "automatic ad generation" business logic, but rather, by introducing specific data structures, prompt statement construction algorithms, generative artificial intelligence model invocation mechanisms, and sentiment analysis and performance feedback loops within the server, it technically improves the computer's processing efficiency, generation accuracy, resource utilization, and communication load in text generation tasks.

[0342] The server runs on one or more computing nodes, which include a central processing unit, a graphics processing unit or other vector computing hardware, main memory, non-volatile memory, and a network interface. The server runs an operating system, a database management system, middleware, and applications on its software. The database management system may be a relational database management system used to manage product information, user information, behavioral history data, sentiment state data, and advertising composition data. The application module adopts a layered architecture, including a data access layer, a feature processing layer, a prompt generation layer, a generative artificial intelligence model invocation layer, a performance analysis layer, and a sentiment analysis integration layer.

[0343] Terminals include smartphones, tablets, personal computers, or other computing devices equipped with display devices, camera devices, and sound acquisition devices. The terminal runs an operating system, browser, or local applications on its software to perform functions such as interface rendering, user interaction acquisition, and multimedia data preprocessing. The terminal interacts with the server via a communication network.

[0344] Users operate the system through a terminal. They input product and user information on the terminal interface and confirm or edit generated advertisements using text data. User actions such as clicking, scrolling, and dwell time while browsing advertisements are collected by the terminal and transmitted to the server. Simultaneously, with the user's consent, the terminal acquires the user's facial expressions and voice data through a camera and audio capture device to infer their emotional state.

[0345] After receiving product and user information from the terminal, the server first converts it into an internal unified data structure in the data access layer. The server represents product information using key-value pairs or structured records, such as storing it as fields like "product category," "price range," and "feature tags." User information is standardized using fields like "age range," "gender," "interest tags," and "purchase preferences." In the feature processing layer, the server uses feature encoding logic to map these fields into vector forms. For example, category fields use one-hot encoding, numerical fields use standardization, and interest tags use multi-hot encoding or embedded vector representation for subsequent prompt construction and performance analysis.

[0346] The server constructs prompts using a combination of rule templates and learnable parameters at the prompt generation layer. The server maintains multiple prompt templates in memory, each containing several placeholders for inserting product information, user information, and emotional state features. The server selects an appropriate template based on predefined non-linear mapping rules and fills in the placeholders according to the current product and user information, thus generating a highly readable natural language prompt. For example, the server can construct the following prompt: "Please create a concise and attractive ad asset text for male users aged 20 to 30 who enjoy gaming and photography, centered around the product '5G gaming phone.' Requirements: Highlight high performance and camera features, use an enthusiastic tone, and keep the text under 30 words." "Users have been frequently searching for 'high-end watches' recently. Please generate an ad asset text that emphasizes quality and design, with a word count of no more than 25 characters." "The user's current sentiment is dissatisfaction. Please generate an advertising asset text based on the 'After-sales Service' commercial material, with a sincere tone to appease the user and explain the improvement measures." By using this structured template and feature binding method, the server can dynamically optimize the prompt statement strategy by adjusting the template and parameters without modifying the underlying code, thereby reducing the system update cost.

[0347] In the generative AI model invocation layer, the server sends the constructed prompts as input to the generative AI model deployed on the computing nodes. The generative AI model can employ a neural network model based on the Transformer architecture. This model includes multi-layer self-attention encoders and decoders, with each layer employing a multi-head attention mechanism, a feedforward fully connected sublayer, and a normalization sublayer. The model's input is a sequence of prompts processed by a tokenizer. Before invoking the model, the server converts the natural language prompts into a tokenized sequence and maps it to a vector sequence using an embedding matrix. When invoking the model, the server sets generation parameters, such as temperature coefficient, sampling strategy (e.g., Top-k sampling or kernel sampling), and maximum generation length, thereby achieving a balance between generation diversity and stability.

[0348] During inference, generative AI models calculate the correlation between the current generation position and historical tags using a self-attention mechanism, and calculate the probability distribution of the next tag based on the trained weight parameters. In the model training phase, the training device pre-trains the model using a large-scale text corpus, employing cross-entropy as the loss function, and updates the model parameters through backpropagation. Data augmentation strategies, such as random masking and synonym replacement, can also be used during training to improve the model's robustness to different prompts. During the inference phase, the server utilizes the pre-trained model parameters without updating the weights, thus rapidly generating advertising text data through forward propagation.

[0349] After receiving the labeled sequence output by the generative artificial intelligence model, the server restores the labeled sequence to natural language text in the text post-processing module, and performs operations such as sensitive word filtering, length truncation, and formatting. Through this rigorous post-processing workflow, the server can reduce redundant and inappropriate content, and improve the quality and controllability of the output text.

[0350] In the ad composition data layer, the server combines the generated ad text data with stored image resources, layout templates, and other elements. The server uses a document object model or structured layout tree to represent the ad composition data, where text nodes correspond to the text data used in the ad, image nodes correspond to image resources, and interactive nodes correspond to elements such as buttons and links. The server can generate different layouts based on different terminal screen sizes to improve the user-side display experience.

[0351] In the performance analysis layer, the server uses historical behavioral data to quantitatively evaluate the performance of the advertising text data. The server reads statistical values ​​related to a specific advertising text data from the database, such as impressions, clicks, and conversions, and calculates performance metrics such as click-through rate, conversion rate, and average dwell time. In the analysis module, the server can use a sliding time window approach to perform time-series analysis on performance metrics and identify trend changes. Simultaneously, the server can use clustering or classification algorithms to categorize user behavior patterns and identify user groups that respond well to specific text versions.

[0352] Based on performance metrics and user segmentation results, the server dynamically adjusts the parameters of the prompt message, such as tone intensity, emphasis on promotional information, and word count range. For example, when the server detects that the click-through rate of a certain version of the text is significantly lower than a preset threshold, it can automatically change the prompt message from "informative" to "call to action," such as changing "introducing product features" to "emphasizing limited-time offers." The server then inputs the updated prompt message back into the generative AI model, thereby achieving adaptive iteration of the copy without human intervention.

[0353] The server processes user state data and emotion-related text data in the sentiment analysis integration layer. The server can call external sentiment recognition services or locally deployed sentiment recognition models to perform facial feature extraction on image data, acoustic feature extraction on audio data, and semantic sentiment classification on text data. The sentiment recognition model can employ a structure combining convolutional neural networks and recurrent neural networks to extract facial expression features from images; alternatively, it can use a Transformer-based speech and text sentiment recognition network. The server inputs these features into the sentiment classifier and outputs sentiment labels (such as joy, dissatisfaction, neutral, etc.) and their corresponding confidence scores.

[0354] After obtaining emotional state data, the server incorporates it as additional conditions into the prompt statements. For example, when the emotional state is joyful, the server adds descriptions such as "excited tone" and "strengthen positive emotional guidance" to the prompt statements; when the emotional state is dissatisfied, it adds constraints such as "sincere tone," "explain improvement measures," and "avoid excessive promotion." This practice of encoding emotional states into prompt statements effectively changes the input distribution of generative AI models, making the text generated by the model more segmented and contextualized under the same product and user attributes, thereby improving the matching degree between content and the user's current psychological state.

[0355] After receiving the advertisement composition data returned by the server, the terminal renders the layout tree locally using its rendering engine, combining text content with image resources for display. At the user interface layer, the terminal records user actions such as scrolling and clicking on advertisements, encoding these events into lightweight log data and periodically sending them to the server. Through this event-level data reporting mechanism, the system can obtain accurate historical behavior records without transmitting large amounts of redundant page data, thereby reducing communication load.

[0356] User participation in the system is no longer limited to one-time input, but rather influences the server's internal prompt construction logic and generative AI model invocation strategy through continuous behavioral and emotional feedback. Because the prompts are designed as structured and adjustable parameterized text instructions, the server can automatically fine-tune the prompts using algorithms, rather than relying on manual modification of text rules. This allows the system to automatically evolve its generation strategy based on statistical results during operation, achieving technically adaptive optimization of the text generation process.

[0357] This invention achieves the following technical effects through the synergy of the aforementioned hardware and software: The server no longer uses static rule concatenation during the advertising text generation process, but instead uses a generative artificial intelligence model combined with structured prompts to generate natural language, significantly improving text diversity and relevance; Through joint analysis of behavioral history data and sentiment state data, the server achieves closed-loop adjustment of prompts, thereby significantly improving the click-through rate and conversion rate of advertising texts in multiple iterations. This performance feedback-based adaptive generation mechanism represents a technical improvement to the computational process and model calling logic; By explicitly constraining the number of words, tone, and emphasis in the prompts, the server reduces unnecessary verbose output during inference by the generative artificial intelligence model, helping to reduce computational load and data transmission volume, and improving the overall system response speed; Through a unified data structure and modular pipeline design, the server internally stores and processes product information, user information, behavioral history data, and sentiment state data in a structured manner, reducing redundant parsing and conversion, and improving data management efficiency.

[0358] Furthermore, the system of this invention can be extended to various implementations. For example, the server can be deployed in a centralized data center or edge computing node, and the terminal can be an in-vehicle terminal, wearable device, etc. The generative artificial intelligence model can be replaced with Transformer networks of different sizes, or a distilled lightweight model can be used to reduce inference latency. The sentiment analysis part can be selected to be based on text only, image only, or voice only, to adapt to different privacy and resource constraints. In these variations, the server always drives the generation and optimization of advertising text data through the interaction between the prompts and the generative artificial intelligence model, and forms a feedback loop through behavioral history data and sentiment state data, thereby maintaining technical improvements in computing performance and generation quality under different hardware and network environments.

[0359] use Figure 14 The processing flow is explained.

[0360] Step 1: The user enters basic information on the terminal. Users enter product and user information in the graphical interface of the terminal and submit it to the server.

[0361] Input: Product information (such as product name, product category, price range, and key features), user information (such as age group, gender, interest tags, and region), and optional targeting objectives (such as click-through rate priority or conversion rate priority).

[0362] Processing: The terminal reads each field from the input control and organizes it into structured data records; the terminal performs local validation on obviously erroneous formats (such as empty fields or illegal characters) and encodes the validated data into a set of key-value pairs.

[0363] Output: The terminal generates a structured request data and sends it to the server as a network request through the communication module.

[0364] Step 2: The terminal sends a request to the server and enters a waiting state. After the user clicks the "Generate Ad Copy" button, the terminal transmits the structured data from step 1 to the server via the network.

[0365] Input: Structured request data (including product information, user information, and delivery targets).

[0366] Processing: The terminal calls the communication interface to serialize the data into a transmission format; the terminal starts a loading animation on the interface and locks the submit button to prevent duplicate submissions.

[0367] Output: The network request data sent to the server, and the "Processing" status displayed on the terminal.

[0368] Step 3: The server receives and stores product information and user information. The server receives network requests from terminals and writes the product and user information into its internal data storage.

[0369] Input: Structured request data received from the terminal.

[0370] Processing: The server parses the request body in the communication module, and validates the field types and required fields; the server writes the validated data into the corresponding table record in the database, cleans the text fields (removing extra spaces and illegal characters), and generates a unique identifier for each record.

[0371] Output: Product information records and user information records stored in the database, along with their corresponding record identifiers.

[0372] Step 4: The server performs feature processing on the input data. The server performs feature encoding on the stored product and user information to prepare for constructing prompt statements and subsequent performance analysis.

[0373] Input: Product information and user information associated with the corresponding record identifier in the database.

[0374] Processing: The server reads records from the database, maps category fields to category indexes, and maps interest tags to tag sets; the server converts these fields into internal feature vectors or feature tag sets according to preset encoding rules, and establishes "product feature objects" and "user feature objects".

[0375] Output: Product feature objects and user feature objects stored in memory.

[0376] Step 5: Server construct prompt statement The server generates prompts for generative artificial intelligence models based on product feature objects and user feature objects.

[0377] Input: Product feature object, user feature object, and target parameters.

[0378] Processing: The server selects a prompt template that matches the current scenario (such as a new product promotion template, discount promotion template, or brand image template), and fills the template placeholders with information such as product name, product category, user age group, gender, interest tags, expected tone, and character limit, generating a prompt statement in natural language.

[0379] Output: One or more natural language prompts, for example: "Please create a concise and attractive ad asset text for male users aged 20 to 30 who enjoy gaming and photography, centered around the product '5G gaming phone.' Requirements: Highlight high performance and camera features, use an enthusiastic tone, and keep the text under 30 words." Step 6: The server calls a generative artificial intelligence model to generate text data for advertising. The server inputs the prompt statement generated in step 5 into the generative artificial intelligence model to generate text data for advertising.

[0380] Input: Natural language prompts, model-generated parameters (such as maximum length, temperature coefficient, sampling strategy).

[0381] Processing: The server segments and tokenizes the prompt statement, converting it into a token sequence; the server inputs the token sequence into a generative artificial intelligence model based on the Transformer structure, performs forward propagation calculations, and progressively predicts the output tokens; the server collects the generated token sequences and restores them to natural language text; the server performs filtering and truncation processing on the generated results, removing fragments that do not meet the length requirements or contain inappropriate content.

[0382] Output: Post-processed text data for advertising, such as: "5G high-speed performance and ultra-clear night shooting make every game an exhilarating experience!" Step 7: Server-side construction of advertising composition data The server embeds the generated advertisement into the advertisement composition data using text data, which is then displayed on the terminal.

[0383] Input: Text data for advertising, product feature objects, and advertising template resources (layout information, image resources) in the system.

[0384] Processing: The server selects a matching ad template (such as a banner, news feed card, or interstitial ad), fills in the text placeholders in the template with the text data for the ad, binds the product image URL or icon resource to the image placeholder, and binds the default button text (such as "View Now") to the interactive element; the server combines these elements into a structured ad data object.

[0385] Output: Ad composition data including text, images, and interactive configurations.

[0386] Step 8: The server returns the advertisement composition data to the terminal. The server sends the advertisement composition data to the terminal via the network, and the terminal renders and previews it based on this data.

[0387] Input: Ad composition data object, target terminal identifier, and session information.

[0388] Processing: The server serializes the advertisement data into a transmission format in the communication module, attaches the necessary identification information, and sends it to the terminal that initiated the request; the server records this generation behavior for subsequent performance statistics.

[0389] Output: The advertisement sent to the terminal constitutes a data response.

[0390] Step 9: The terminal renders and displays the ad preview. The terminal receives the advertising composition data returned by the server and displays the generated advertisement on the local interface.

[0391] Input: Ad composition data.

[0392] Processing: The terminal parses the advertisement composition data, loads text content and image resources, and arranges controls such as text boxes, image areas, and buttons on the interface according to the template layout information; the terminal calls the rendering interface of the display device to draw the combined advertisement content on the screen.

[0393] Output: An advertisement preview displayed on the terminal display device.

[0394] Step 10: Users can review and edit advertising copy on the terminal. Users can view the server-generated advertising copy on their devices and choose to accept, partially modify, or regenerate it.

[0395] Input: Advertising copy and interface editing controls displayed on the terminal.

[0396] Processing: The user modifies the text data of the advertisement in the text editing box, or clicks the "Regenerate" button; the terminal reads the modified text or records the user's "Regenerate" operation, and packages the request according to the operation type: for modification operation, the modified text is sent to the server as new copy; for regeneration operation, only the regeneration instruction, original product information and user information identifier are sent.

[0397] Output: Edited result data sent to the server or regenerated request data.

[0398] Step 11: Terminal collects and reports user behavior information The terminal collects operation and browsing information while the user is browsing advertisements and sends it to the server periodically.

[0399] Input: The user's actual actions on the ad interface (clicks, swipes, dwell time, etc.).

[0400] Processing: The terminal captures user interaction events in the interface event listener, adds a timestamp and advertising identifier to each event, and encodes the events as behavior log entries; after reaching a certain number or time interval, the terminal packages the behavior logs in batches and reports them to the server over the network.

[0401] Output: Batch of historical behavior data sent to the server.

[0402] Step 12: The server records and organizes historical behavioral data. The server receives historical behavioral data reported by the terminal, and organizes and stores it.

[0403] Input: Batch of behavior logs reported by the terminal (including events such as clicks, displays, and dwell times).

[0404] Processing: The server parses the behavior logs and categorizes events by ad identifier, user identifier, and timeline; the server writes the events to the behavior history database table and standardizes some fields (such as timestamps) to create indexes for statistical analysis.

[0405] Output: Structured records of behavioral history data stored in the database.

[0406] Step 13: The server calculates performance metrics based on historical behavioral data. The server performs statistical analysis on the stored historical behavioral data to obtain performance metrics for the text data used in advertising.

[0407] Input: Historical behavioral data associated with a specific advertisement.

[0408] Processing: The server uses a statistics module to aggregate and calculate fields such as number of impressions, number of clicks, and number of conversions to obtain metrics such as click-through rate, conversion rate, and average dwell time; the server can use a sliding time window to calculate metrics in segments over time to observe trends over time.

[0409] Output: A set of performance metrics corresponding to the text data for each advertisement.

[0410] Step 14: Server adjusts prompts based on performance metrics The server uses performance metrics to optimize the conditions for constructing subsequent prompt statements.

[0411] Input: a set of performance metrics, the current prompt statement template, and parameter settings.

[0412] Processing: The server compares the performance metrics of a certain advertisement's text data with a preset threshold or historical average level. When it finds that the click-through rate is low or the conversion rate is declining, the server modifies the prompt statement strategy according to predefined rules, such as adding a "limited-time offer" description, shortening the number of words, and strengthening the call to action. The server writes the updated strategy into the prompt statement construction parameters for use in the next round of prompt statement generation.

[0413] Output: Adjusted prompt generation parameters and updated prompt template configuration.

[0414] Step 15: The terminal collects user status data and uploads it for sentiment analysis. After obtaining user authorization, the terminal uses a camera and a sound acquisition device to collect the user's facial expressions and voice data, and then uploads them to the server.

[0415] Input: Facial image frames and audio clips of the user while watching the advertisement.

[0416] Processing: The terminal captures image frames at fixed intervals and performs resolution scaling and compression on the images; the terminal samples and encodes the voice signal; the terminal packages the preprocessed multimedia data into user state data and sends it to the server via a secure connection.

[0417] Output: Batch of user status data sent to the server.

[0418] Step 16: The server performs sentiment analysis and generates sentiment state data. The server performs sentiment analysis on the received user status data to obtain sentiment status data.

[0419] Input: User status data (image data, voice data, etc.) from the terminal.

[0420] Processing: The server inputs image data into the emotion recognition model to extract facial key points and expression features; it inputs speech data into the acoustic feature extraction module to extract features such as pitch, energy, and speech rate; the server combines these features and inputs them into the emotion classifier to perform forward inference and output the emotion category (such as joy, dissatisfaction, neutral) and confidence score; the server filters unstable results based on the confidence score and retains only reliable emotion labels.

[0421] Output: Sentiment state data records associated with the user and advertising context.

[0422] Step 17: The server generates emotionally sensitive prompts based on emotional state. The server combines emotional state data with product feature objects and user feature objects to generate prompt statements that include emotional constraints.

[0423] Input: sentiment state data, product feature objects, user feature objects, and current prompt statement strategy parameters.

[0424] Processing: When the emotion is joy, the server adds descriptions such as "excited tone" and "enhancing positive experience" to the prompt; when the emotion is dissatisfaction, it adds constraints such as "sincere tone," "explaining improvement measures," and "avoiding excessive sales pitches." The server fills these emotional constraints into the template along with the original product and user information, and placeholders include emotional tone and whether to add compensation information.

[0425] Output: Emotionally sensitive prompts, for example: "The user's current sentiment is dissatisfaction. Please generate an advertising asset text based on the 'After-sales Service' commercial material, with a sincere tone to appease the user and explain the improvement measures." Step 18: The server regenerates the advertising text data based on emotionally sensitive prompts. The server uses the emotionally sensitive prompts generated in step 17 to call a generative artificial intelligence model to generate advertising text data that is adapted to the emotional state.

[0426] Input: Emotionally sensitive prompts and model generation parameters.

[0427] Processing: The server repeats the prompt statement tokenization and model forward inference process, but because sentiment constraint information is added to the input, the model tends to select expressions that match the current sentiment in the probability space; the server filters the generated results again to ensure that the tone is consistent with the sentiment constraint.

[0428] Output: Advertising text data that is appropriate for the current emotional state, such as: "To improve your experience, we are speeding up our services and offering exclusive offers, and we look forward to regaining your trust." Step 19: The server updates the ad composition data in real time and pushes it to the terminal. The server replaces the original text with new text data for the advertisement, updates the advertisement composition data, and pushes it to the terminal in real time.

[0429] Input: Text data for ads that are tailored to emotional states, and the current ad as a data object.

[0430] Processing: The server locates the text node in the ad composition data, replaces the original content with the new text, updates the version number and timestamp of the composition data, and sends the updated ad composition data to the terminal through a fast push channel or polling response.

[0431] Output: Updated ad composition data, and an update notification sent to the terminal.

[0432] Step 20: The terminal refreshes the displayed content and continues to collect subsequent data. After receiving the updated ad composition data, the terminal refreshes the interface to display the new ad copy and continues to collect user behavior and status data.

[0433] Input: Updated ad composition data.

[0434] Processing: The terminal stops displaying the old text and re-renders the new text and layout to the display device; the terminal continues to listen for new user clicks, pauses, and other behaviors, and continues to collect images and audio when permitted, providing data for subsequent server iterations.

[0435] Output: The latest advertising content presented to users, as well as historical behavioral data and user status data that will be reported subsequently.

[0436] The specific processing unit 290 sends the result of the specific processing to the smart device 14. In the smart device 14, the control unit 46A causes the output device 40 to output the result of the specific processing. The microphone 38B acquires sound representing user input regarding the result of the specific processing. The control unit 46A sends the sound data representing user input acquired by the microphone 38B to the data processing device 12. In the data processing device 12, the specific processing unit 290 acquires the sound data.

[0437] Data generation model 58 is a so-called generative AI (Artificial Intelligence). Examples of data generation models 58 include ChatGPT (registered trademark) (accessible via the internet (URL: https: / / openai.com / blog / chatgpt)). Data generation model 58 is obtained through deep learning on a neural network. Input to data generation model 58 are prompt words containing instructions, and inference data such as sound data representing sound, text data representing text, and image data representing images (e.g., still image data or animation data). Data generation model 58 infers from the input inference data based on the instructions represented by the prompt words and outputs the inference result in one or more data forms, such as sound data, text data, and image data. Data generation model 58 includes, for example, text generation AI, image generation AI, and multimodal generation AI. Here, inference refers to, for example, analysis, classification, prediction, and / or induction. The specific processing unit 290 performs the aforementioned specific processing while using data generation model 58. The data generation model 58 can also be a model finely tuned to output inference results from prompts that do not contain instructions. In this case, the data generation model 58 can output inference results based on prompts that do not contain instructions. The data processing apparatus 12, etc., includes various data generation models 58, including AI other than the generation AI. AI other than the generation AI can be, for example, linear regression, logistic regression, decision trees, random forests, support vector machines (SVM), k-means clustering, convolutional neural networks (CNN), recurrent neural networks (RNN), generative adversarial networks (GAN), or Naive Bayes, and can perform various processes, but is not limited to this example. Furthermore, the AI ​​can also be an AI agent. Furthermore, when the processing of the above-mentioned parts is performed by AI, the processing can be performed partially or entirely by AI, but is not limited to this example. Furthermore, the processing performed by the AI ​​including the generation AI can be replaced by processing in the rule base, and the processing in the rule base can also be replaced by processing performed by the AI ​​including the generation AI.

[0438] Furthermore, the processing of the aforementioned data processing system 10 is performed by the specific processing unit 290 of the data processing device 12 or the control unit 46A of the smart device 14, but it can also be performed by both the specific processing unit 290 of the data processing device 12 and the control unit 46A of the smart device 14. Additionally, the specific processing unit 290 of the data processing device 12 acquires or collects information required for processing from the smart device 14 or external devices, and the smart device 14 acquires or collects information required for processing from the data processing device 12 or external devices.

[0439] For example, the collection unit is implemented by the control unit 46A of the smart device 14 or the specific processing unit 290 of the data processing device 12. For example, the acquisition unit uses the camera 42 or communication I / F 44 of the smart device 14 to acquire step data, which is then processed by the specific processing unit 290 of the data processing device 12. For example, the analysis unit is implemented by the specific processing unit 290 of the data processing device 12, which analyzes the data from the collection unit and the acquisition unit. For example, the generation unit is implemented by the specific processing unit 290 of the data processing device 12, which uses a generation AI to generate a menu. For example, the serving unit is implemented by the output device 40 of the smart device 14 or the specific processing unit 290 of the data processing device 12, which provides the generated menu to the user. The correspondence between each unit and the device or control unit is not limited to the above examples and various changes can be made.

[0440] In the above embodiments, examples of specific processing by the data processing device 12 are given, but the technology disclosed herein is not limited to this, and specific processing may also be performed by the smart device 14.

[0441] Second Implementation Method Figure 3 An example of the configuration of the data processing system 210 according to the second embodiment is shown.

[0442] like Figure 3 As shown, the data processing system 210 includes a data processing device 12 and smart glasses 214. A server can be cited as an example of the data processing device 12.

[0443] The data processing apparatus 12 includes a computer 22, a database 24, and a communication I / F 26. The computer 22 is an example of a "computer" as understood in this disclosure. The computer 22 includes a processor 28, RAM 30, and memory 32. The processor 28, RAM 30, and memory 32 are connected to a bus 34. Furthermore, the database 24 and the communication I / F 26 are also connected to the bus 34. The communication I / F 26 is connected to a network 54. Examples of the network 54 include a WAN (Wide Area Network) and / or a LAN (Local Area Network).

[0444] The smart glasses 214 include a computer 36, a microphone 238, a speaker 240, a camera 42, and a communication I / F 44. The computer 36 includes a processor 46, RAM 48, and memory 50. The processor 46, RAM 48, and memory 50 are connected to a bus 52. Furthermore, the microphone 238, speaker 240, camera 42, and communication I / F 44 are also connected to the bus 52.

[0445] Microphone 238 receives instructions from user 20 by receiving sounds emitted by user 20. Microphone 238 captures sounds emitted by user 20 and converts the captured sounds into sound data, which is then output to processor 46. Speaker 240 outputs sound according to instructions from processor 46.

[0446] Camera 42 is a small digital camera equipped with an optical system such as a lens, aperture and shutter, and imaging elements such as a CMOS (Complementary Metal-Oxide-Semiconductor) image sensor or a CCD (Charge Coupled Device) image sensor, which captures images of the user 20's surroundings (e.g., the field of view defined by an angle equivalent to the field of vision of an average healthy person).

[0447] Communication I / F44 is connected to network 54. Communication I / F44 and 26 are responsible for the transmission and reception of various information between processor 46 and processor 28 via network 54. The transmission and reception of various information between processor 46 and processor 28 using communication I / F44 and 26 is performed in a secure state.

[0448] Figure 4 This illustrates an example of the main functions of the data processing device 12 and the smart glasses 214. For example... Figure 4 As shown, in the data processing device 12, specific processing is performed by the processor 28. The specific processing program 56 is stored in the memory 32.

[0449] The specific processing program 56 is an example of a "program" involved in the technology of this disclosure. The processor 28 reads the specific processing program 56 from the memory 32 and executes the read specific processing program 56 on the RAM 30. Specific processing is implemented by the processor 28 operating as a specific processing unit 290 according to the specific processing program 56 executed on the RAM 30.

[0450] The memory 32 stores a data generation model 58 and an emotion-specific model 59. The data generation model 58 and the emotion-specific model 59 are used by the specific processing unit 290. The specific processing unit 290 can use the emotion-specific model 59 to infer the user's emotions and perform specific processing based on the user's emotions. In the emotion inference function (emotion-specific function) using the emotion-specific model 59, various inferences and predictions related to the user's emotions are performed, including inferences and predictions of the user's emotions, but this is not limited to this example. Furthermore, emotion inference and prediction may also include, for example, emotion analysis (parsing).

[0451] In the smart glasses 214, the processor 46 performs reception and output processing. The memory 50 stores the reception and output program 60. The processor 46 reads the reception and output program 60 from the memory 50 and executes the read reception and output program 60 on the RAM 48. The reception and output processing is implemented by the processor 46 operating as a control unit 46A according to the reception and output program 60 executed on the RAM 48. Furthermore, the smart glasses 214 has the same data generation model and emotion-specific model as the data generation model 58 and the emotion-specific model 59, and these models can also be used to perform the same processing as the specific processing unit 290.

[0452] Next, the specific processing of the specific processing unit 290 of the data processing device 12 will be described. Each part of the system described below is implemented by the data processing device 12 and the smart glasses 214. In the following description, the data processing device 12 will be referred to as the "server" and the smart glasses 214 will be referred to as the "terminal".

[0453] Example 1 The process is the same as that of the specific process described in Embodiment 1 in the first embodiment above, so the description is omitted.

[0454] Application Example 1 The process is the same as that in the specific processing described in Application Example 1 of the first embodiment above, so the description is omitted.

[0455] Example 2 The process is the same as that of the specific process in Embodiment 2 described in the first embodiment above, so the description is omitted.

[0456] Application Example 2 The process is the same as that in the specific processing described in Application Example 2 of the first embodiment above, so the description is omitted.

[0457] The specific processing unit 290 sends the result of the specific processing to the smart glasses 214. In the smart glasses 214, the control unit 46A outputs the result of the specific processing to the speaker 240. The microphone 238 acquires sound input representing the user's input regarding the result of the specific processing. The control unit 46A sends the sound data representing the user's input acquired by the microphone 238 to the data processing device 12. In the data processing device 12, the specific processing unit 290 acquires the sound data.

[0458] Data generation model 58 is a so-called generative AI (Artificial Intelligence). Examples of data generation models 58 include ChatGPT (registered trademark) (accessible via the internet (URL: https: / / openai.com / blog / chatgpt)). Data generation model 58 is obtained through deep learning on a neural network. Input to data generation model 58 are prompt words containing instructions, and inference data such as sound data representing sound, text data representing text, and image data representing images (e.g., still image data or animation data). Data generation model 58 infers from the input inference data based on the instructions represented by the prompt words and outputs the inference result in one or more data forms, such as sound data, text data, and image data. Data generation model 58 includes, for example, text generation AI, image generation AI, and multimodal generation AI. Here, inference refers to, for example, analysis, classification, prediction, and / or induction. The specific processing unit 290 performs the aforementioned specific processing while using data generation model 58. The data generation model 58 can also be a model finely tuned to output inference results from prompts that do not contain instructions. In this case, the data generation model 58 can output inference results based on prompts that do not contain instructions. The data processing apparatus 12, etc., includes various data generation models 58, including AI other than the generation AI. AI other than the generation AI can be, for example, linear regression, logistic regression, decision trees, random forests, support vector machines (SVM), k-means clustering, convolutional neural networks (CNN), recurrent neural networks (RNN), generative adversarial networks (GAN), or Naive Bayes, and can perform various processes, but is not limited to this example. Furthermore, the AI ​​can also be an AI agent. Furthermore, when the processing of the above-mentioned parts is performed by AI, the processing can be performed partially or entirely by AI, but is not limited to this example. Furthermore, the processing performed by the AI ​​including the generation AI can be replaced by processing in the rule base, and the processing in the rule base can also be replaced by processing performed by the AI ​​including the generation AI.

[0459] Furthermore, the processing of the aforementioned data processing system 10 is performed by the specific processing unit 290 of the data processing device 12 or the control unit 46A of the smart glasses 214, but it can also be performed by both the specific processing unit 290 of the data processing device 12 and the control unit 46A of the smart glasses 214. Additionally, the specific processing unit 290 of the data processing device 12 acquires or collects information required for processing from the smart glasses 214 or external devices, and the smart glasses 214 acquires or collects information required for processing from the data processing device 12 or external devices.

[0460] For example, the collection unit is implemented by the control unit 46A of the smart glasses 214 or the specific processing unit 290 of the data processing device 12. For example, the acquisition unit uses the camera 42 or communication I / F 44 of the smart glasses 214 to acquire step data, which is then processed by the specific processing unit 290 of the data processing device 12. For example, the analysis unit is implemented by the specific processing unit 290 of the data processing device 12, which analyzes the data from the collection unit and the acquisition unit. For example, the generation unit is implemented by the specific processing unit 290 of the data processing device 12, which uses a generation AI to generate a menu. For example, the serving unit is implemented by the speaker 240 of the smart glasses 214 or the specific processing unit 290 of the data processing device 12, which provides the generated menu to the user. The correspondence between each unit and the device or control unit is not limited to the above examples and various changes can be made.

[0461] In the above embodiments, examples of specific processing by the data processing device 12 are given, but the technology disclosed herein is not limited to this, and specific processing may also be performed by the smart glasses 214.

[0462] Third Implementation Method Figure 5 An example of the configuration of the data processing system 310 according to the third embodiment is shown.

[0463] like Figure 5 As shown, the data processing system 310 includes a data processing device 12 and a head-mounted terminal 314. A server can be cited as an example of the data processing device 12.

[0464] The data processing apparatus 12 includes a computer 22, a database 24, and a communication I / F 26. The computer 22 is an example of a "computer" as understood in this disclosure. The computer 22 includes a processor 28, RAM 30, and memory 32. The processor 28, RAM 30, and memory 32 are connected to a bus 34. Furthermore, the database 24 and the communication I / F 26 are also connected to the bus 34. The communication I / F 26 is connected to a network 54. Examples of the network 54 include a WAN (Wide Area Network) and / or a LAN (Local Area Network).

[0465] The head-mounted terminal 314 includes a computer 36, a microphone 238, a speaker 240, a camera 42, a communication I / F 44, and a display 343. The computer 36 includes a processor 46, RAM 48, and memory 50. The processor 46, RAM 48, and memory 50 are connected to a bus 52. Furthermore, the microphone 238, speaker 240, camera 42, display 343, and communication I / F 44 are also connected to the bus 52.

[0466] Microphone 238 receives instructions from user 20 by receiving sounds emitted by user 20. Microphone 238 captures sounds emitted by user 20 and converts the captured sounds into sound data, which is then output to processor 46. Speaker 240 outputs sound according to instructions from processor 46.

[0467] Camera 42 is a small digital camera equipped with an optical system such as a lens, aperture and shutter, and imaging elements such as a CMOS (Complementary Metal-Oxide-Semiconductor) image sensor or a CCD (Charge Coupled Device) image sensor, which captures images of the user 20's surroundings (e.g., the field of view defined by an angle equivalent to the field of vision of an average healthy person).

[0468] Communication I / F44 is connected to network 54. Communication I / F44 and 26 are responsible for the transmission and reception of various information between processor 46 and processor 28 via network 54. The transmission and reception of various information between processor 46 and processor 28 using communication I / F44 and 26 is performed in a secure state.

[0469] Figure 6 This illustrates an example of the main functions of the data processing device 12 and the head-mounted terminal 314. For example... Figure 6 As shown, in the data processing device 12, specific processing is performed by the processor 28. The specific processing program 56 is stored in the memory 32.

[0470] The specific processing program 56 is an example of a "program" involved in the technology of this disclosure. The processor 28 reads the specific processing program 56 from the memory 32 and executes the read specific processing program 56 on the RAM 30. Specific processing is implemented by the processor 28 operating as a specific processing unit 290 according to the specific processing program 56 executed on the RAM 30.

[0471] The memory 32 stores the data generation model 58 and the emotion-specific model 59. The data generation model 58 and the emotion-specific model 59 are used by the specific processing unit 290.

[0472] In the head-mounted terminal 314, the processor 46 performs the acceptance / output processing. The memory 50 stores the acceptance / output program 60. The processor 46 reads the acceptance / output program 60 from the memory 50 and executes the read acceptance / output program 60 on the RAM 48. The acceptance / output processing is implemented by the processor 46 operating as a control unit 46A according to the acceptance / output program 60 executed on the RAM 48.

[0473] Next, the specific processing of the specific processing unit 290 of the data processing device 12 will be described. Each part of the system described below is implemented by the data processing device 12 and the head-mounted terminal 314. In the following description, the data processing device 12 will be referred to as the "server" and the head-mounted terminal 314 will be referred to as the "terminal".

[0474] Example 1 The process is the same as that of the specific process described in Embodiment 1 in the first embodiment above, so the description is omitted.

[0475] Application Example 1 The process is the same as that in the specific processing described in Application Example 1 of the first embodiment above, so the description is omitted.

[0476] Example 2 The process is the same as that of the specific process in Embodiment 2 described in the first embodiment above, so the description is omitted.

[0477] Application Example 2 The process is the same as that in the specific processing described in Application Example 2 of the first embodiment above, so the description is omitted.

[0478] The specific processing unit 290 sends the result of the specific processing to the head-mounted terminal 314. In the head-mounted terminal 314, the control unit 46A outputs the result of the specific processing to the speaker 240 and the display 343. The microphone 238 acquires sound input representing the user's input regarding the result of the specific processing. The control unit 46A sends the sound data representing the user's input acquired by the microphone 238 to the data processing device 12. In the data processing device 12, the specific processing unit 290 acquires the sound data.

[0479] Data generation model 58 is a so-called generative AI (Artificial Intelligence). Examples of data generation models 58 include ChatGPT (registered trademark) (accessible via the internet (URL: https: / / openai.com / blog / chatgpt)). Data generation model 58 is obtained through deep learning on a neural network. Input to data generation model 58 includes prompt words containing instructions, and inference data such as sound data representing sound, text data representing text, and image data representing images (e.g., still image data or animation data). Data generation model 58 infers the input inference data based on the instructions represented by the prompt words and outputs the inference result in one or more data forms such as sound data, text data, and image data. Data generation model 58 includes, for example, text generation AI, image generation AI, and multimodal generation AI. Here, inference refers to, for example, analysis, classification, prediction, and / or induction. The specific processing unit 290 performs the aforementioned specific processing while using data generation model 58. The data generation model 58 can also be a model finely tuned to output inference results from prompts that do not contain instructions. In this case, the data generation model 58 can output inference results based on prompts that do not contain instructions. The data processing apparatus 12, etc., includes various data generation models 58, including AI other than the generation AI. AI other than the generation AI can be, for example, linear regression, logistic regression, decision trees, random forests, support vector machines (SVM), k-means clustering, convolutional neural networks (CNN), recurrent neural networks (RNN), generative adversarial networks (GAN), or Naive Bayes, and can perform various processes, but is not limited to this example. Furthermore, the AI ​​can also be an AI agent. Furthermore, when the processing of the above-mentioned parts is performed by AI, the processing can be performed partially or entirely by AI, but is not limited to this example. Furthermore, the processing performed by the AI ​​including the generation AI can be replaced by processing in the rule base, and the processing in the rule base can also be replaced by processing performed by the AI ​​including the generation AI.

[0480] Furthermore, the processing of the aforementioned data processing system 10 is performed by the specific processing unit 290 of the data processing device 12 or the control unit 46A of the head-mounted terminal 314, but it can also be performed by both the specific processing unit 290 of the data processing device 12 and the control unit 46A of the head-mounted terminal 314. Additionally, the specific processing unit 290 of the data processing device 12 acquires or collects information required for processing from the head-mounted terminal 314 or external devices, and the head-mounted terminal 314 acquires or collects information required for processing from the data processing device 12 or external devices.

[0481] For example, the collection unit is implemented by the control unit 46A of the head-mounted terminal 314 or the specific processing unit 290 of the data processing device 12. For example, the acquisition unit uses the camera 42 or communication I / F 44 of the head-mounted terminal 314 to acquire step data, which is then processed by the specific processing unit 290 of the data processing device 12. For example, the analysis unit is implemented by the specific processing unit 290 of the data processing device 12, which analyzes the data from the collection unit and the acquisition unit. For example, the generation unit is implemented by the specific processing unit 290 of the data processing device 12, which uses a generation AI to generate a menu. For example, the serving unit is implemented by the speaker 240 and display 343 of the head-mounted terminal 314 or the specific processing unit 290 of the data processing device 12, which provides the generated menu to the user. The correspondence between each unit and the device or control unit is not limited to the above examples and various changes can be made.

[0482] In the above embodiments, examples of specific processing by the data processing device 12 are given, but the technology disclosed herein is not limited to this, and specific processing may also be performed by the head-mounted terminal 314.

[0483] Fourth Implementation Method Figure 7 An example of the configuration of the data processing system 410 according to the fourth embodiment is shown.

[0484] like Figure 7 As shown, the data processing system 410 includes a data processing device 12 and a robot 414. A server can be cited as an example of the data processing device 12.

[0485] The data processing apparatus 12 includes a computer 22, a database 24, and a communication I / F 26. The computer 22 is an example of a "computer" as understood in this disclosure. The computer 22 includes a processor 28, RAM 30, and memory 32. The processor 28, RAM 30, and memory 32 are connected to a bus 34. Furthermore, the database 24 and the communication I / F 26 are also connected to the bus 34. The communication I / F 26 is connected to a network 54. Examples of the network 54 include a WAN (Wide Area Network) and / or a LAN (Local Area Network).

[0486] Robot 414 includes a computer 36, a microphone 238, a speaker 240, a camera 42, a communication I / F 44, and a controlled object 443. The computer 36 includes a processor 46, RAM 48, and memory 50. The processor 46, RAM 48, and memory 50 are connected to a bus 52. Furthermore, the microphone 238, speaker 240, camera 42, controlled object 443, and communication I / F 44 are also connected to the bus 52.

[0487] Microphone 238 receives instructions from user 20 by receiving sounds emitted by user 20. Microphone 238 captures sounds emitted by user 20 and converts the captured sounds into sound data, which is then output to processor 46. Speaker 240 outputs sound according to instructions from processor 46.

[0488] Camera 42 is a small digital camera equipped with an optical system such as a lens, aperture and shutter, and imaging elements such as a CMOS (Complementary Metal-Oxide-Semiconductor) image sensor or a CCD (Charge Coupled Device) image sensor, to photograph the area around robot 414 (e.g., the field of view defined by a perspective equivalent to the field of vision of an average healthy person).

[0489] Communication I / F44 is connected to network 54. Communication I / F44 and 26 are responsible for the transmission and reception of various information between processor 46 and processor 28 via network 54. The transmission and reception of various information between processor 46 and processor 28 using communication I / F44 and 26 is performed in a secure state.

[0490] The controlled object 443 includes a display device, LEDs (light-emitting diodes) for the eyes, and motors for driving the arms, hands, and feet. The posture or movement of the robot 414 is controlled by controlling the motors in the arms, hands, and feet. Some of the robot 414's emotions can be expressed by controlling these motors. In addition, the robot 414's facial expressions can also be expressed by controlling the illumination state of the LEDs in its eyes.

[0491] Figure 8 This illustrates an example of the main functions of the data processing device 12 and the robot 414. For example... Figure 8 As shown, in the data processing device 12, specific processing is performed by the processor 28. The specific processing program 56 is stored in the memory 32.

[0492] The specific processing program 56 is an example of a "program" involved in the technology of this disclosure. The processor 28 reads the specific processing program 56 from the memory 32 and executes the read specific processing program 56 on the RAM 30. Specific processing is implemented by the processor 28 operating as a specific processing unit 290 according to the specific processing program 56 executed on the RAM 30.

[0493] The memory 32 stores the data generation model 58 and the emotion-specific model 59. The data generation model 58 and the emotion-specific model 59 are used by the specific processing unit 290.

[0494] In robot 414, the processor 46 performs the acceptance and output processing. The memory 50 stores the acceptance and output program 60. The processor 46 reads the acceptance and output program 60 from the memory 50 and executes the read acceptance and output program 60 on RAM 48. The acceptance and output processing is implemented by the processor 46 acting as the control unit 46A according to the acceptance and output program 60 executed on RAM 48.

[0495] Next, the specific processing of the specific processing unit 290 of the data processing device 12 will be described. Each part of the system described below is implemented by the data processing device 12 and the robot 414. In the following description, the data processing device 12 will be referred to as the "server" and the robot 414 will be referred to as the "terminal".

[0496] Example 1 The process is the same as that of the specific process described in Embodiment 1 in the first embodiment above, so the description is omitted.

[0497] Application Example 1 The process is the same as that in the specific processing described in Application Example 1 of the first embodiment above, so the description is omitted.

[0498] Example 2 The process is the same as that of the specific process in Embodiment 2 described in the first embodiment above, so the description is omitted.

[0499] Application Example 2 The process is the same as that in the specific processing described in Application Example 2 of the first embodiment above, so the description is omitted.

[0500] The specific processing unit 290 sends the result of the specific processing to the robot 414. In the robot 414, the control unit 46A outputs the result of the specific processing to the speaker 240 and the controlled object 443. The microphone 238 acquires sound input representing the result of the specific processing. The control unit 46A sends the sound data representing the user input acquired by the microphone 238 to the data processing device 12. In the data processing device 12, the specific processing unit 290 acquires the sound data.

[0501] Data generation model 58 is a so-called generative AI (Artificial Intelligence). Examples of data generation models 58 include ChatGPT (registered trademark) (accessible via the internet (URL: https: / / openai.com / blog / chatgpt)). Data generation model 58 is obtained through deep learning on a neural network. Input to data generation model 58 are prompt words containing instructions, and inference data such as sound data representing sound, text data representing text, and image data representing images (e.g., still image data or animation data). Data generation model 58 infers from the input inference data based on the instructions represented by the prompt words and outputs the inference result in one or more data forms, such as sound data, text data, and image data. Data generation model 58 includes, for example, text generation AI, image generation AI, and multimodal generation AI. Here, inference refers to, for example, analysis, classification, prediction, and / or induction. The specific processing unit 290 performs the aforementioned specific processing while using data generation model 58. The data generation model 58 can also be a model finely tuned to output inference results from prompts that do not contain instructions. In this case, the data generation model 58 can output inference results based on prompts that do not contain instructions. The data processing apparatus 12, etc., includes various data generation models 58, including AI other than the generation AI. AI other than the generation AI can be, for example, linear regression, logistic regression, decision trees, random forests, support vector machines (SVM), k-means clustering, convolutional neural networks (CNN), recurrent neural networks (RNN), generative adversarial networks (GAN), or Naive Bayes, and can perform various processes, but is not limited to this example. Furthermore, the AI ​​can also be an AI agent. Furthermore, when the processing of the above-mentioned parts is performed by AI, the processing can be performed partially or entirely by AI, but is not limited to this example. Furthermore, the processing performed by the AI ​​including the generation AI can be replaced by processing in the rule base, and the processing in the rule base can also be replaced by processing performed by the AI ​​including the generation AI.

[0502] Furthermore, the processing of the aforementioned data processing system 10 is performed by the specific processing unit 290 of the data processing device 12 or the control unit 46A of the robot 414, but it can also be performed by both the specific processing unit 290 of the data processing device 12 and the control unit 46A of the robot 414. Additionally, the specific processing unit 290 of the data processing device 12 acquires or collects information required for processing from the robot 414 or external devices, and the robot 414 acquires or collects information required for processing from the data processing device 12 or external devices.

[0503] For example, the collection unit is implemented by the control unit 46A of the robot 414 or the specific processing unit 290 of the data processing device 12. For example, the acquisition unit uses the camera 42 or communication I / F 44 of the robot 414 to acquire step data, which is then processed by the specific processing unit 290 of the data processing device 12. For example, the analysis unit is implemented by the specific processing unit 290 of the data processing device 12, which analyzes the data from the collection unit and the acquisition unit. For example, the generation unit is implemented by the specific processing unit 290 of the data processing device 12, which uses a generation AI to generate a menu. For example, the serving unit is implemented by the speaker 240 of the robot 414 and the control object 443 or the specific processing unit 290 of the data processing device 12, which provides the generated menu to the user. The correspondence between each unit and the device or control unit is not limited to the above examples and various changes can be made.

[0504] In the above embodiments, examples of specific processing by the data processing device 12 are given, but the technology disclosed herein is not limited to this, and specific processing may also be performed by the robot 414.

[0505] Furthermore, the emotion-specific model 59, acting as an emotion engine, can determine a user's emotion based on a specific mapping. Specifically, the emotion-specific model 59 can determine a user's emotion based on an emotion graph that serves as a specific mapping (see [reference]). Figure 9 The emotion-specific model 59 can also determine the robot's emotion, and the specific processing unit 290 performs specific processing based on the robot's emotions.

[0506] Figure 9 This is a diagram representing an emotion map 400 that maps multiple emotions. In the emotion map 400, emotions are arranged in concentric circles radiating from the center. The closer to the center of the concentric circles, the more primitive the emotion is. On the outer side of the concentric circles, emotions representing states or behaviors arising from mood are arranged. Emotions are concepts that include feelings and mental states. Emotions generated by reactions occurring in the brain are arranged roughly to the left of the concentric circles. Emotions derived from situational judgments are arranged roughly to the right of the concentric circles. Emotions generated by reactions occurring in the brain and derived from situational judgments are arranged roughly above and below the concentric circles. Furthermore, "pleasant" emotions are arranged above the concentric circles, and "unpleasant" emotions are arranged below them. Thus, in the emotion map 400, multiple emotions are mapped based on the structure that generates emotions, and emotions that are likely to occur simultaneously are mapped close to each other.

[0507] These emotions are distributed at the three o'clock position of the emotion map 400, typically fluctuating between peace and anxiety. In the right half of the emotion map 400, situational awareness dominates over internal sensation, thus resulting in an impression of calm.

[0508] The inner side of the emotion map 400 represents the inner state, while the outer side represents behavior. Therefore, the further outward you are from the emotion map 400, the more visible the emotion becomes (manifested in behavior).

[0509] Here, human emotions are based on various balances such as posture and blood sugar levels. When these balances deviate from an ideal state, it indicates an unpleasant state; when they approach the ideal state, it indicates a pleasant state. Emotions in robots, cars, motorcycles, etc., can also be created in the following way: based on various balances such as posture and remaining battery power, when these balances deviate from an ideal state, it indicates an unpleasant state; when they approach the ideal state, it indicates a pleasant state. Emotion maps can be generated, for example, based on Dr. Mitsuyoshi's emotion map (Research on a Brain Physiological Signal Analysis System for Voice Emotion Recognition and Emotion, Tokushima University, Doctoral Dissertation: https: / / ci.nii.ac.jp / naid / 500000375379). In the left half of the emotion map, emotions belonging to the sensory-dominated region, called "response," are arranged. Furthermore, in the right half of the emotion map, emotions belonging to the situational cognition-dominated region, called "situation," are arranged.

[0510] In the emotion map, two types of emotions that promote learning are defined. One is a negative emotion on the situational side, in the middle or peripheral region of "repentance" or "reflection." This occurs when the robot experiences negative emotions such as "I don't want to experience this feeling again" or "I don't want to be blamed again." The other is a positive emotion on the response side, near the "desire" region. This occurs when there are positive feelings such as "wanting more" or "wanting to know more."

[0511] The emotion-specific model 59 inputs user input into a pre-trained neural network to obtain emotion values ​​representing each emotion shown in the emotion map 400, thereby determining the user's emotion. This neural network is pre-trained based on multiple training data sets, which are combinations of user input and emotion values ​​representing each emotion shown in the emotion map 400. Furthermore, this neural network... Figure 10 As shown in the sentiment graph 900, it was trained in a way that sentiments that are configured close to each other have similar values. Figure 10 The text shows examples of emotions such as "peace of mind", "stability", and "reassurance" that have similar emotion values.

[0512] The above description focuses on the functions of the data processing device 12, but the system of this disclosure is not necessarily installed on a server. The system of this disclosure can also be installed as a general information processing system. This disclosure can also be installed, for example, as a software program running on a personal computer, an application running on a smartphone, etc. The method of this disclosure can also be provided to users in the form of SaaS (Software as a Service).

[0513] In the above embodiments, an example of a specific process being performed by a single computer 22 is given. However, the technology disclosed herein is not limited to this, and the specific process can also be distributed among multiple computers, including computer 22. For example, the data generation model 58 can be located on an external device of the data processing apparatus 12, where data is generated based on the input data.

[0514] In the above embodiments, examples of storing a specific processing program 56 in the memory 32 have been described, but the technology disclosed herein is not limited thereto. For example, the specific processing program 56 may also be stored in a portable computer-readable non-transitory storage medium such as a USB (Universal Serial Bus) memory. The specific processing program 56 stored in the non-transitory storage medium is installed into the computer 22 of the data processing device 12. The processor 28 executes specific processing according to the specific processing program 56.

[0515] Alternatively, a specific processing program 56 may be pre-stored in a storage device such as a server connected to the data processing device 12 via the network 54, and the specific processing program 56 may be downloaded and installed on the computer 22 according to the requirements of the data processing device 12.

[0516] In addition, it is not necessary to store all the specific processing program 56 in the storage device such as the server connected to the data processing device 12 via the network 54 or in the memory 32; a portion of the specific processing program 56 may be stored in advance.

[0517] As hardware resources for performing specific processes, various processors, as shown below, can be used. For example, a CPU can be listed as a processor, which functions as a general-purpose processor that performs specific processes by executing software, i.e., a program. Furthermore, processors can be listed as special-purpose circuits such as FPGAs (Field-Programmable Gate Arrays), PLDs (Programmable Logic Devices), or ASICs (Application-Specific Integrated Circuits), which are processors with circuitry specifically designed to perform specific processes. Each processor has built-in or connected memory, and each processor executes specific processes using that memory.

[0518] The hardware resources for performing a specific process can consist of one of these various processors, or a combination of two or more processors of the same or different types (e.g., a combination of multiple FPGAs, or a combination of a CPU and an FPGA). Alternatively, the hardware resources for performing a specific process can be a single processor.

[0519] As an example of a single processor, there are two approaches: First, a processor is composed of a combination of one or more CPUs and software, which functions as a hardware resource to perform a specific process; second, as represented by a SoC (System-on-a-chip), a processor is used to implement the functionality of the entire system, which includes multiple hardware resources for performing a specific process, using a single IC (Integrated Circuit) chip. In this way, the specific process is implemented by using one or more of the aforementioned processors as hardware resources.

[0520] Furthermore, the hardware architecture of these various processors, more specifically, can utilize circuits that combine semiconductor elements and other circuit components. Moreover, the specific process described above is just one example. Therefore, without departing from the main point, unnecessary steps can certainly be deleted, new steps added, or the processing order changed.

[0521] The descriptions and illustrations above are detailed explanations of a portion of the technology disclosed herein, and are merely one example of the technology disclosed herein. For example, the above descriptions of the structure, function, effect, and results are just one example of the structure, function, effect, and results of a portion of the technology disclosed herein. Therefore, without departing from the spirit of the technology disclosed herein, unnecessary parts may be deleted, new elements added, or replacements may be made to the descriptions and illustrations above. Furthermore, to avoid confusion and facilitate understanding of a portion of the technology disclosed herein, explanations of common technical knowledge that do not require special explanation under the premise of being able to implement the technology disclosed herein have been omitted from the descriptions and illustrations above.

[0522] All documents, patent applications and technical specifications set forth in this specification are incorporated herein by reference to the same extent that each document, patent application and technical specification is specifically and individually described therein and referenced by reference.

[0523] In addition, the following notes are provided in response to the above explanation.

[0524] Example 1 (Note 1) An information processing system, characterized in that it comprises: An apparatus for receiving product characteristic information and user attribute information from a terminal, and storing the product characteristic information and user attribute information as structured data; A device for performing information extraction and normalization processing on the structured data to generate element information containing product information elements and target user information elements; An apparatus for assembling prompt statements for a generative artificial intelligence model according to a preset template, based on the element information and generation condition information including style conditions, word count conditions and output quantity conditions. An apparatus for sending the prompt statement and generation parameter information via a network to an information processing device that runs the generative artificial intelligence model, and for performing natural language generation processing through the generative artificial intelligence model to obtain asset text containing advertising text candidates; An apparatus for performing post-processing on the acquired asset text, including prohibited word determination, word count determination, and format sorting, thereby generating advertising asset data; An apparatus for embedding the advertising asset data into an advertising delivery data structure to generate advertising display data for a display terminal; A device for sending the advertising display data to the terminal and for the terminal to present the advertising display data to the user in a visual manner.

[0525] (Note 2) The information processing system according to Appendix 1 is characterized in that, The system includes a device for associating and storing the product characteristic information, the user attribute information, and the user behavior history information as historical information, and a device for updating the prompt statement or the generation condition information based on the historical information, thereby optimizing the asset text generated by the generative artificial intelligence model.

[0526] (Note 3) The information processing system according to Appendix 1 is characterized in that, The system includes a device for continuously parsing user behavior information obtained from the terminal, and a device for changing the element information or the generation condition information in real time according to the parsing results, and using the changed prompt statement to perform the regeneration processing of the generative artificial intelligence model, thereby updating the asset text in real time.

[0527] Application Example 1 (Note 1) An information processing system, characterized in that it comprises: A unit used to acquire and preprocess attribute information related to products and attribute information related to users; A unit for generating natural language representations including user profiles and product profiles based on the preprocessed user-related attribute information and the preprocessed product-related attribute information, and for specifying the input conditions, output conditions, and word count constraints of the generative artificial intelligence model. A unit used to input the prompt statement into the generative artificial intelligence model and automatically generate asset text composed of text elements for advertising through deep learning processing; A unit for performing prohibited expression detection, information consistency verification, and length adjustment on the automatically generated asset text, so as to filter and correct the content of the asset text; A unit for embedding the filtered and corrected asset text into advertising display data and sending the advertising display data to a terminal device via a communication unit; A unit for parsing the advertising display data in the terminal device and displaying an advertisement containing the asset text on the user's display device in real time; A unit for acquiring user behavior information regarding browsing, selecting, or operating on the advertisement in the terminal device, and sending the behavior information to the system; A unit for updating user-related attribute information and interest level information based on the sent behavioral information, and applying the updated information to the subsequent generation process of the prompt statement and the asset text.

[0528] (Note 2) The information processing system according to Appendix 1 is characterized in that, It also includes a unit for extracting feature quantities from the product-related attribute information and the user-related attribute information, calculating the relevance of multiple product candidates based on the feature quantities, and selecting a product candidate as the object of the prompt statement generation according to the relevance.

[0529] (Note 3) The information processing system according to Appendix 1 is characterized in that, It also includes a unit for continuously executing the process of inputting the prompt statement into the generative artificial intelligence model to generate asset text based on the acquisition results of the user behavior information, and resending the advertising display data corresponding to the asset text, which is updated in real time with user behavior, to the terminal device.

[0530] Example 2 (Note 1) An information processing system, characterized in that it comprises: A device for obtaining input information, including information characteristics and user attributes, by a processing unit in an information processing apparatus; An apparatus for the processing unit to construct a prompt statement as input to a generative artificial intelligence model based on the input information and advertising delivery conditions, and to generate the prompt statement. An apparatus for inputting the prompt statement into the generative artificial intelligence model by the processing unit to automatically generate advertising text information, and for obtaining the generated advertising text information; An apparatus for sending the acquired advertising text information from the processing unit to the advertising management device, and for the advertising management device to automatically insert the advertising text information into the content area of ​​the advertisement to update the advertising information; An apparatus for obtaining updated advertising information from the processing unit and displaying the updated advertising information to a user terminal via a display device; An apparatus for obtaining, through the user terminal, the user's confirmation result or correction instruction for the advertising content as acceptance information by the processing unit, and for repeatedly executing the construction of the prompt statement and the generation and processing of advertising text information based on input information containing the acceptance information.

[0531] (Note 2) According to the information processing system described in Appendix 1, the processing unit is configured to acquire accumulated information based on user behavior history and response information to advertisements, generate feature quantities representing information characteristics and user attributes based on the accumulated information, and improve the content of the advertisement text information generated using the generative artificial intelligence model by appending the feature quantities to the prompt statement.

[0532] (Note 3) According to the information processing system described in Appendix 1, the processing unit is configured to acquire changes in state information, including user input information, user attributes, and behavioral history, at a predetermined period or when an event occurs, dynamically update the prompt statement based on the acquired state information, and perform real-time processing of advertising text information generation using the generative artificial intelligence model and updating of the advertising information.

[0533] Application Example 2 (Note 1) An information processing system, characterized in that it comprises: A device for acquiring product information and user information as input data; Apparatus for constructing prompt statements for input into a generative artificial intelligence model based on the product information and the user information; A device for inputting the prompt statement into the generative artificial intelligence model to automatically generate text data for advertising; A means for embedding the generated text data of the advertisement into the advertisement composition data; A device for outputting the advertising composition data to a display device to present advertising information; A device for collecting user operation and browsing information as behavioral history data; An apparatus for calculating performance metrics of the advertising text data based on the behavioral history data, changing the composition conditions of the prompt statement according to the calculation results, and inputting the prompt statement back into the generative artificial intelligence model to update the generated advertising text data; A device for acquiring user status data through an image acquisition device and a sound acquisition device and performing sentiment analysis processing to determine the sentiment status data; An apparatus for generating a prompt statement corresponding to the emotional state based on the determined emotional state data, the product information, and the user information; inputting the prompt statement into the generative artificial intelligence model to automatically generate advertising text data adapted to the emotional state; and using the advertising text data to update the advertising composition data in real time.

[0534] (Note 2) The information processing system according to Appendix 1 is characterized in that it further includes: An apparatus for parsing evaluation and response information obtained from users, processing it as emotion-related text data, determining the user's emotional state data based on the emotion-related text data, adjusting the content and expression of the prompt statement according to the emotional state data to generate the advertising text data, and applying the advertising text data to the advertising composition data.

[0535] (Note 3) The information processing system according to Appendix 1 is characterized in that it further includes: An apparatus for sequentially acquiring the behavioral history data and the emotional state data over time, automatically reconstructing the prompt statements input to the generative artificial intelligence model each time they are acquired, and reflecting the text data of the advertisement generated by the generative artificial intelligence model into the advertisement composition data to continuously update the advertisement information being displayed.

Claims

1. An information processing system, characterized in that, include: processor; The processor is configured to: receive material characteristics and user attributes as input; generate prompt text based on the input information to instruct a generative artificial intelligence model to generate asset text; and automatically generate asset text based on the prompt text. The generated asset text is applied to the advertisement; and the display of the advertisement is controlled.

2. The information processing system according to claim 1, characterized in that, The processor is also configured to optimize the asset text based on the user's behavioral history.

3. The information processing system according to claim 1, characterized in that, The processor is also configured to update the asset text in real time based on the user's attributes and behavior.

Images