Method and system for data generation
The data generation method and system address the challenge of diverse content handling by generating high-quality data sets using AI models, ensuring consistency and accuracy across different content types, thereby enhancing AI model performance.
Patent Information
- Application Number
- PCT/KR2026/005927
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2026-05-07
- Filing Date
- 2026-05-08
- Publication Date
- 2026-11-12
AI Technical Summary
Existing data generation technologies for artificial intelligence models, particularly large language models, struggle to handle diverse content types effectively, leading to insufficient data quality and performance in creative writing applications.
A data generation method and system that utilizes an artificial intelligence model to process first data and attribute data, generating high-quality data sets by considering structural and stylistic dimensions of various content types, with quality evaluation and selection based on preset criteria.
Enables the generation of high-quality data sets that accurately reflect diverse content types, ensuring consistency, accuracy, and reliability, improving the performance of AI models in handling varied content.
Smart Images

Figure KR2026005927_12112026_PF_FP_ABST
Abstract
Description
Data generation methods and systems
[0001] The present invention relates to a data generation method and system, and more specifically, to a data generation method and system that generates high-quality data.
[0002] Artificial intelligence can be defined as a technology that realizes human learning, reasoning, perceptual, and natural language understanding abilities into computer programs. This artificial intelligence has achieved continuous development through deep learning.
[0003] In particular, driven by the advancement of artificial intelligence, various language models have been developed. These models have reached a level where they not only recognize text and understand its meaning but also extract and classify information from vast amounts of text-based data, such as documents, and even generate text directly.
[0004] These language models are actively utilized in various fields and exist in diverse areas where text-based tasks can be performed, such as search engines, document creation (e.g., resume writing, report writing, posting, etc.), free conversation on various topics, data parsing from given text (e.g., data summarization, classification, etc.), provision of expertise, programming, and converting given sentences into sentences of an appropriate style.
[0005] In this regard, Large Language Models (LLMs) have recently emerged, which understand and generate human language through prior training on vast amounts of text data. Unlike traditional chatbots, which are manually built and provide only limited answers, LLMs are demonstrating innovation in the artificial intelligence market by showcasing technological capabilities that provide conversational responses to user queries that are comparable to those of humans.
[0006] Currently, these large-scale language models are being utilized in various fields. For example, large-scale language model-based approaches are used for creative writing, and this accounts for a significant portion of actual user interactions.
[0007] Accordingly, various studies are being conducted to build high-quality training data for creative writing. While these studies have advanced data-driven modeling, evaluation, and alignment techniques, they have often overlooked the unique structural and functional constraints of other content types (or genres) that are substantially different by equating creative writing with story generation. In other words, there was a limitation in that they could not adequately handle diverse content.
[0008] To address these limitations, there is a need for new data generation technologies that encompass diverse content and can build high-quality data to improve the performance of artificial intelligence models.
[0009] The present invention is intended to provide a data generation method and system capable of automatically generating high-quality data based on artificial intelligence.
[0010] More specifically, the present invention aims to provide a data generation method and system capable of generating high-quality data for improving the performance of an artificial intelligence model while encompassing various content.
[0011] In particular, the present invention aims to provide a data generation method and system capable of generating high-quality data specialized for various content based on artificial intelligence.
[0012] Furthermore, the present invention aims to solve the problem of insufficient data and to provide a data generation method and system capable of automatically generating and providing data encompassing various content.
[0013] To solve the problem described above, a data generation method according to the present invention, performed by a computer, presents data processing that generates high-quality data corresponding to various content using an artificial intelligence model. More specifically, the data generation method according to the present invention may include the steps of: receiving at least one first data corresponding to a first content; receiving type information for a second content different from the first content and generating at least one attribute data related to the second content based on the type information; generating at least one second data corresponding to the type information based on an artificial intelligence model using input data including at least one of the first data and the attribute data; and providing a data set related to the type information using the second data.
[0014] In an embodiment, in the step of generating the second data, at least one of the first data and the attribute data can be input into the artificial intelligence model to generate the second data corresponding to the type information.
[0015] In an embodiment, the method may include the steps of configuring a prompt using at least one of the first data, the type information, and the attribute data, inputting the prompt into the artificial intelligence model, and obtaining the second data generated from the artificial intelligence model based on the prompt.
[0016] In an embodiment, the first data comprises first query-response pairs having at least one first query corresponding to type information for the first content and a response to the first query, and the first query-response pairs may be included as few-shot examples in the prompt.
[0017] In an embodiment, the step of generating the second data may include inputting the prompt into the artificial intelligence model to generate at least one second query corresponding to the type information for the second content and generating a response to the second query.
[0018] In an embodiment, the step of generating the second data generates the second data using the second query and the response to the second query, and the second data may include second query-response pairs configured to have the second query and the response to the second query.
[0019] In an embodiment, the step of performing a quality evaluation of the second data is further included, and in the step of providing the data set, the data set including at least a portion of the second data may be provided based on the quality evaluation.
[0020] In an embodiment, the step of performing the quality evaluation may include the step of performing a quality evaluation of the second query-response pairs using an evaluation model, and the step of selecting at least some of the second query-response pairs that satisfy a preset criterion based on the quality evaluation.
[0021] In the embodiment, in the selection step, a score based on the quality evaluation of each of the second query-response pairs is calculated using the evaluation model, and based on the calculated score, at least some of the second query-response pairs that satisfy the preset criteria can be selected.
[0022] In an embodiment, the step of providing the data set may provide the data set including at least some of the pairs selected based on the quality evaluation.
[0023] In an embodiment, the first data is input through a user terminal, and the data set can be provided through the user terminal.
[0024] In an embodiment, the result of generating the data set can be visually provided through a user interface.
[0025] In an embodiment, in the step of generating the attribute data, the attribute data related to the second content may be generated based on at least some of the attributes defined for the type information for the second content among the attributes defined for each of the type information for each of the multiple different contents.
[0026] In an embodiment, in the step of generating the attribute data, at least some of the attributes defined for the type information for the second content among the attributes may be selected, and the attribute data related to the second content may be generated based on the selected at least some of the attributes.
[0027] In the embodiment, the attributes may represent structural and stylistic dimensions that distinguish the plurality of contents.
[0028] In an embodiment, the data set may include data corresponding to each of the type information.
[0029] In the embodiment, the data may include query-response pairs corresponding to each of the type information.
[0030] Additionally, the data generation method of the present invention may include the steps of: receiving first data corresponding to at least one piece of content; receiving type information corresponding to a type different from the content and generating attribute data that reflects the structural characteristics of the different type based on the type information; generating input data for a generation model using at least one of the first data and the attribute data, and generating second data corresponding to the type information from the generation model based on the input data; and generating a data set of content related to the type information using the second data.
[0031] A data generation system according to the present invention, comprising a memory configured to store executable instructions and one or more processors configured to perform operations by executing one or more instructions, wherein the memory stores an artificial intelligence model configured to generate data based on sets of attributes defined for a plurality of different content types, and the at least one processor receives at least one first data corresponding to a first content, specifies a content type for a second content corresponding to a target content, generates at least one attribute data related to the second content based on a set of attributes defined for the second content among the sets of attributes, generates structured input information to be provided as input to the artificial intelligence model using at least one of the first data and the attribute data, and inputs the structured input information to the artificial intelligence model to receive at least one second data corresponding to the content type output from the artificial intelligence model based on the input data, and the system can provide a data set related to the content type based on the second data received from the artificial intelligence model.
[0032] A program according to the present invention is a program that is executed by one or more processors in an electronic device and can be stored on a computer-readable recording medium, comprising: a first instruction for receiving first data corresponding to at least one content; a second instruction for receiving type information for a second content different from the first content and generating at least one attribute data related to the second content based on an attribute set defined for the second content among attribute sets defined for a plurality of different content types; a third instruction for generating at least one second data corresponding to the type information based on an artificial intelligence model using input data including at least one of the first data and the attribute data; and a fourth instruction for generating a data set related to the type information using the second data, wherein the attributes included in the attribute set may represent structural and stylistic dimensions that distinguish a plurality of different content.
[0033] As described above, according to the data generation method and system of the present invention, by utilizing first data corresponding to the first content and attribute data related to the content to be generated, second data having a different type from the first content can be generated based on an artificial intelligence model. That is, the present invention enables the generation and expansion of data of various content types and can construct a data set encompassing various content types. Accordingly, as the difference between the type information for the first content and the type information for the second content is reflected in the attribute data generation step, the present invention allows a data set including multiple different content types to be constructed using the artificial intelligence model. Accordingly, in the present invention, a conversion from an existing content type to data corresponding to a content type different from the existing content type can be performed, and through such conversion, the data set can be utilized in various application fields.
[0034] Furthermore, according to the data generation method and system of the present invention, a prompt can be configured using at least one of the first data, type information, and attribute data, and the prompt can be input into an artificial intelligence model to generate second data. That is, the present invention can generate high-quality data that is structural and consistent for various content types, and can simultaneously ensure accuracy and diversity in data generation.
[0035] Furthermore, according to the data generation method and system of the present invention, attribute data is generated based on structural and stylistic attributes defined for each type of content information, and new data is generated using this, thereby enabling the generation of customized data that reflects the characteristics of each content type. Through this, the present invention can construct a data set including various content types.
[0036] Furthermore, the present invention can provide a dataset containing high-quality data by using an evaluation model to perform a quality evaluation of the second dataset and selecting at least a portion of the second dataset based on the quality evaluation. Accordingly, the reliability and usability of the provided dataset can be improved, and the quality of the data used for training or services can be stably maintained.
[0037] Furthermore, according to the data generation method and system of the present invention, by providing the generated data set through a user terminal, it supports the user in intuitively checking and utilizing data of various content types, and can improve the overall efficiency of data generation and utilization and user convenience.
[0038] FIG. 1 illustrates an example of a block diagram of a computing system in which the present invention can be implemented.
[0039] FIG. 2 illustrates an example of a block diagram of a computing device that may be included in a user computing device, a server computing system, and a training computing system, as an embodiment of a computing system in which the present invention can be implemented.
[0040] Figure 3 illustrates an example of a block diagram from another perspective of a computing device, which is one of the components of a computing system.
[0041] FIGS. 4 and FIGS. 5 are conceptual diagrams for explaining a data generation system according to the present invention.
[0042] FIG. 6 is a flowchart illustrating a data generation method according to the present invention.
[0043] FIGS. 7, FIGS. 8, FIGS. 9, FIGS. 10, FIGS. 11 and FIGS. 12 are conceptual diagrams for explaining a data generation method according to the present invention.
[0044] FIGS. 13 and FIGS. 14 are drawings illustrating mathematical expressions for explaining a data generation method according to the present invention.
[0045] FIGS. 15 and 16 are drawings illustrating performance related to a data generation method and system according to the present invention.
[0046] Hereinafter, embodiments disclosed in this specification will be described in detail with reference to the attached drawings. Identical or similar components are assigned the same reference number regardless of the drawing symbols, and redundant descriptions thereof will be omitted. The suffixes "module" and "part" used for components in the following description are assigned or used interchangeably solely for the ease of drafting the specification and do not inherently possess distinct meanings or roles. Furthermore, in describing embodiments disclosed in this specification, if it is determined that a detailed description of related prior art could obscure the essence of the embodiments disclosed in this specification, such detailed description will be omitted. Additionally, the attached drawings are intended only to facilitate understanding of the embodiments disclosed in this specification; the technical concept disclosed in this specification is not limited by the attached drawings, and it should be understood that they include all modifications, equivalents, and substitutions that fall within the spirit and technical scope of the present invention.
[0047] Terms including ordinal numbers, such as first, second, etc., may be used to describe various components, but said components are not limited by said terms. These terms are used solely for the purpose of distinguishing one component from another.
[0048] When it is stated that one component is "connected" or "connected" to another component, it should be understood that while it may be directly connected or connected to that other component, there may also be other components in between. On the other hand, when it is stated that one component is "directly connected" or "directly connected" to another component, it should be understood that there are no other components in between.
[0049] A singular expression includes a plural expression unless the context clearly indicates otherwise.
[0050] In this application, terms such as “comprising” or “having” are intended to specify the existence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.
[0051] Hereinafter, the meanings and equivalent expressions of key terms used throughout this specification are defined. The following definitions may serve as intrinsic evidence for the interpretation of terms described in the claims, and even if other terms are used in this specification, the equivalent expressions listed below may be understood to mean the same concept.
[0052] In this specification, "first data" refers to data corresponding to the first content and means a data object capable of constituting at least a part of the input provided to an artificial intelligence model, and may include at least one of a query, a query-response pair, a sample included in a seed data set, reference data, and data obtained from external storage. "First data" may also be referred to in this specification as "input data," "source data," "seed data," "reference data," and "existing content data," and all such expressions may be understood to mean the same concept as "first data" as used in this specification.
[0053] In this specification, "second data" refers to a data object generated from an artificial intelligence model to correspond to a content type different from that of the first data, and may include at least one of a generated query, a generated response, a generated query-response pair, and a combination thereof. "Second data" may also be referred to in this specification as "generated data," "output data," "response data," "synthesized data," and "target content data," and all such expressions may be understood to mean the same concept as "second data" as used in this specification.
[0054] In this specification, "attribute data" refers to data representing at least one of structural characteristics and stylistic characteristics that distinguish a specific content type from other content types, and may include at least one of a value, description, or explanation corresponding to at least some of the attributes among the set of attributes defined for each content type. "Attribute data" may also be referred to in this specification as "attribute information," "dimensional descriptor," "style descriptor," "genre attribute," and "type-specific attribute," and these expressions may be understood to mean the same concept.
[0055] Additionally, in this specification, "structure" may also be expressed as "structural characteristic," "structural aspect," or "structural dimension," and "style" may also be expressed as "stylish characteristic," "stylish aspect," or "stylish dimension," and these expressions may be understood to mean the same concept.
[0056] In this specification, "type information" refers to information identifying the category or genre of content, and may include at least one of a text label, an identifier, an index, a class value, and a combination thereof. The term "type information" is used as a technical term to specify the direction of creation of an artificial intelligence model, distinct from unstructured data in the input stage, and is not intended to cause unnecessary duplication or indefiniteness. In this specification, "type information" may also be referred to as "content type information," "genre information," "genre label," "content category," and "target classification," and these expressions are understood to mean the same concept.
[0057] In this specification, "structured input information" refers to data input to an artificial intelligence model, wherein at least one of the first data, the type information, and the attribute data is arranged according to a pre-set template or format. "Structured input information" may also be referred to in this specification as "prompt," "input prompt," "structured prompt," "template instance," and "formatted input representation," and these expressions may be understood to mean the same concept.
[0058] In this specification, "artificial intelligence model" refers to a machine learning-based model trained to generate the second data in response to the input of the structured input information, and may include at least one of a large-scale language model (LLM), a generative AI model, a transformer-based model, a deep neural network, and a combination thereof. In this specification, "artificial intelligence model" may also be referred to as a "generative model," a "generative AI model," a "training target model," a "language model-based generator," and an "inference model," and these expressions may be understood to mean the same concept.
[0059] In this specification, "evaluation model" refers to a machine learning-based model trained to perform a quality evaluation of the second data and calculate an evaluation score, and may be composed of at least one of the same model as the artificial intelligence model or another distinct model. In this specification, "evaluation model" may also be referred to as an "evaluator model," "scoring model," "judge model," "quality classifier," and "rubric-based evaluator," and these expressions may be understood to mean the same concept.
[0060] In this specification, "data set" refers to a data object in which a plurality of data samples are aggregated according to a pre-established structure, and the data set generated in the present invention may be configured to include a plurality of query-response pairs corresponding to each of a plurality of different content types. In this specification, "data set" may also be referred to as "training data set," "corpus," "data set," "sample pool," and "final data set," and these expressions may be understood to mean the same concept.
[0061] Terms such as “comprising” or “having” as used in this specification indicate the existence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not excluding in advance the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.
[0062] Meanwhile, in the field of constructing multi-genre creative writing data, the generation of training data specialized for each different content type has primarily relied on the subjective judgment of human experts, such as writers, editors, or annotators. For example, traditionally, training data was constructed through a process in which experts directly identified the structural and stylistic characteristics required for each content type—such as poetry, lyrics, movie scripts, rap, and TV series—and manually drafted queries or responses reflecting those characteristics; however, the consistency of the results could vary depending on the worker's proficiency, experience, and subjective aesthetic judgment.
[0063] The data generation system according to the present invention can automatically perform the above task based on a set of attributes defined for each content type (e.g., story attributes, poetry attributes, lyrics attributes, movie script attributes, rap attributes, character design attributes, game design attributes, TV series attributes, etc. exemplified in Table 1). Specifically, the data generation system according to the present invention, which includes, as a first rule, identification of a set of attributes corresponding to a target content type; as a second rule, sampling of a subset from said set of attributes (condition of number k of attributes); as a third rule, configuration of structured input information including said subset and few-shot examples; and as a fourth rule, calculation of a score based on an evaluation model for generated query-response pairs, and pre-configured selection according to said criteria based on a set of attributes defined by content type (e.g., story attributes, poetry attributes, lyrics attributes, movie script attributes, rap attributes, character design attributes, game design attributes, TV series attributes, etc. exemplified in Table 1), can derive said data set by applying according to said application order based on a set of attributes defined by content type (e.g., story attributes, poetry attributes, lyrics attributes, movie script attributes, rap attributes, character design attributes, game design attributes, TV series attributes, etc. exemplified in Table 1). Such rule-based automation can replace conventional subjective judgment with a structured decision procedure.
[0064] Quality control can be achieved by an evaluation model-based screening rule through the above automation rule by means of "a step of generating at least one attribute data related to the second content based on type information" and "a step of generating at least one second data corresponding to the type information based on an artificial intelligence model using input data including at least one of the first data and the attribute data."
[0065] The data generation process according to the present invention is characterized by data scale and computational complexity that are difficult to perform sequentially by a single human entity. For example, in one embodiment of the present invention, multiple attributes (e.g., 5 to 15) are defined per content type, and as a combination of subset sampling (number of attributes k=0 to maximum value) and few-shot examples is applied to multiple content types (e.g., 13 or more content types listed in Table 1), the number of cases for attribute subsets corresponding to a single input prompt can increase combinatorially with respect to the number of content types and the number of attributes. Furthermore, since the calculation of scores by an evaluation model for each generated query-response pair, the application of statistical criteria based on the overall distribution (e.g., deviation from the mean, percentile, quartile, range based on variance or standard deviation), and screening according to preset criteria are additionally performed, the data processing of the present invention may have characteristics that are difficult to perform in reality by the cognitive processing of a single entity in terms of required throughput and consistency.
[0066] The above data processing structure may be implemented by the explicit limitation of the “artificial intelligence model” and “evaluation model” of the present invention and “at least one processor connected to memory and configured to perform data processing based on the artificial intelligence model and the evaluation model.”
[0067] Meanwhile, FIG. 1 illustrates an example of a block diagram of a computing system in which the present invention can be implemented. In this regard, the data generation system according to the present invention can be implemented through a computing device described below and can perform data processing related to the data generation method described in this specification.
[0068] Referring to FIG. 1, a computing system (10000) that performs a method for generating high-quality data corresponding to type information for various content according to one embodiment of the present invention may include at least one computing device. At this time, the at least one computing device may be a single processor or a multi-processor computing device.
[0069] The components of at least one computing device of the present invention may include various hardware components such as one or more processors, memory, other hardware, and a system bus (not shown) that connects various system components so that they can transmit and receive data to and from each other (e.g., telecommutatively connected, physically connected, electrically connected), and the components of at least one computing device are not limited thereto and may be diverse.
[0070] Meanwhile, at least one computing device included in a computing system (10000) that performs a method for generating high-quality data corresponding to type information for various content may be connected to communicate via a network (1070). For example, at least one computing device included in the computing system (10000) may be clustered or may be part of a local area network (LAN). Additionally, at least one computing device may be part of a wide area network (WAN) or connected to at least one of a client-server network and a peer-to-peer network within the cloud.
[0071] Meanwhile, when at least one computing device is used in at least one of a network environment and a cloud computing environment, the at least one computing device may be connected to at least one of a public and a private network through a network interface or an adapter. In one embodiment, other communication connection devices, such as a modem, may be used to establish communication through the network. The modem may be at least one of an internal modem and an external modem, and may be connected to a system bus through a network interface or a specific mechanism, etc. A wireless network component consisting of an interface and an antenna may be coupled to the network through a device such as an access point, a peer computer, etc. In the present invention, the method by which at least one computing device is connected to communicate through the network (1070) is not limited and may be connected to communicate in a manner different from the described examples.
[0072] Furthermore, other computer-type devices or systems not shown in FIG. 1 may also interact technically with at least one computing device or other system through one or more connections to the network (1070) via a network interface. Here, the network interface may include network interface equipment such as a physical network interface controller (NIC) or a virtual network interface (VIF).
[0073] The network (1070) of the present invention may include various forms such as the Internet, WLAN (Wireless LAN), Wi-Fi (Wireless-Fidelity), Wi-Fi Direct, DLNA (Digital Living Network Alliance), WiBro (Wireless Broadband), WiMAX (World Interoperability for Microwave Access), HSDPA (High Speed Downlink Packet Access), HSUPA (High Speed Uplink Packet Access), LTE (Long Term Evolution), LTE-A (Long Term Evolution-Advanced), 5G (5th Generation Mobile Telecommunication), Bluetooth (Bluetooth™), RFID (Radio Frequency Identification), Infrared Data Association (IrDA), UWB (Ultra-Wideband), ZigBee, NFC (Near Field Communication), Wireless USB (Wireless Universal Serial Bus), etc., and in the present invention, data transmission may be performed based on standard communication protocols such as TCP / IP, HTTP, SSL, etc.
[0074] A computing system (10000) that performs a method for generating high-quality data corresponding to type information for various content according to the present invention may include at least one of a user computing device (1010), a training computing system (1050), and a server computing system (1030).
[0075] A user computing device (1010) according to the present invention may be understood as a computing device comprising at least one processor (1011) and a memory (1012) that perform a method of generating high-quality data corresponding to type information for various content. For example, the user computing device (1010) may include at least one computing device among a smartphone, a smart TV, a laptop computer, a desktop computer, a digital broadcasting terminal, a PDA (personal digital assistants), a PMP (portable multimedia player), a navigation device, a slate PC, a tablet PC, an ultrabook, a wearable device (e.g., a smartwatch, a smart glass, and a head-mounted display).
[0076] At least one processor (1011) constituting the user computing device (1010) may include at least one general-purpose processor and at least one special-purpose processor. For example, at least one processor (1011) constituting the user computing device (1010) may be composed of at least one or a plurality of electrically connected processors among a central processing unit (CPU), a graphics processing unit (GPU), a tensor processing unit (TPU), a neural network processing unit (NPU), an arithmetic logic unit (ALU), a floating-point arithmetic unit (FPU), an application integrated circuit, an application semiconductor (ASIC), digital signal processing devices (DSPs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, and electrical units for performing micro and other functions.
[0077] Furthermore, at least one processor (1011) may be configured to execute a computer-readable instruction contained in memory (1012) or other instructions described in the present invention.
[0078] The memory (1012) constituting the user computing device (1010) according to the present invention may include volatile memory, non-volatile memory, fixed media, removable media, magnetic media, optical media, semiconductor media, and other types of physically durable storage media.
[0079] For example, memory (1012) may include one or more non-transient / transient computer-readable storage media such as RAM, ROM, HDD (Hard Disk Drive), SSD (Solid State Disk), SSD (Silicon Disk Drive), EEPROM, EPROM, flash memory device, magnetic disk, etc., and may include web storage of a server that performs the storage function of memory on the internet. Such memory (1012) may store data and instructions necessary for at least one processor (1011) to perform the operation of an application for generating high-quality data corresponding to type information for various content.
[0080] A user computing device (1010) may include one or more user input components (1021) that detect user input. For example, the user input component (1021) may also be referred to as a user interface module. The user input component (1021) may include a touch screen, a computer mouse, a keyboard, a keypad, a touchpad, a trackball, a joystick, a voice recognition module, or other similar devices. However, the present invention does not limit the type of user input component (1021). In this case, the user input component (1021) in the present invention does not necessarily mean a hardware means, but can be understood as a channel for receiving input from a user. Meanwhile, the user of the present invention may refer to an automated agent, script, playback software, etc., that operates on behalf of one or more people.
[0081] A user can interact with a computing system (10000) including at least one computing device through input text, touch, voice, movement, computer vision, gesture, or other forms of input and output using a user input component (1021). For example, the user input component (1021) may include at least one of a command line interface (CLI), a graphical user interface (GUI), a natural user interface (NUI), a voice command interface, and other user interface (UI) representations.
[0082] Between the user input component (1021) and the user computing device (1010), one or more application programming interface (API) calls may be made based on user input received from a user interface or network.
[0083] Here, the expression "based on" may be interpreted to include cases where it is based on the use of a specific configuration, modified from, derived from, influenced by, dependent on, or otherwise derived from a specific configuration. In some embodiments, an API call may be configured for a specific API, which may be interpreted or converted into an API call configured for another API. Here, an API may refer to a defined interface or connection between computers or between computer programs.
[0084] In one embodiment, the user computing device (1010) may store at least one machine learning model (1020). For example, the user computing device (1010) may be various machine learning models, such as a plurality of neural networks (e.g., deep neural networks) that perform a method of generating high-quality data corresponding to type information for various content, or other types of machine learning models including non-linear models or linear models, and may be composed of a combination thereof.
[0085] According to an embodiment of the present invention, a user computing device (1010) may perform a method of generating high-quality data corresponding to type information for various content by using a local or external machine learning model (1020). Alternatively, the user computing device (1010) may perform a method of generating high-quality data corresponding to type information for various content by using a machine learning model (1040) provided by a server.
[0086] In addition, according to another embodiment of the present invention, a server computing system (1030) communicating with a user computing device (1010) can provide data specialized in type information for various content to the user computing device (1010) via an application or the web in accordance with a request from a user received through the user computing device (1010).
[0087] In addition, according to another embodiment of the present invention, by performing a method of generating high-quality data corresponding to type information for various content by linking at least a part of a user computing device (1010) and a server computing system (1030) with each other, data specialized for type information for various content can be provided to the user.
[0088] Additionally, according to various embodiments of the present invention, a user computing device (1010) and a server computing system (1030) can learn machine learning models (1020, 1040) performed in a method of generating high-quality data corresponding to type information for various content through interaction with a training computing system (1050) that is communicatedly connected via a network (1070). In this case, the training computing system (1050) may be a computing system separate from the server computing system (1030). Alternatively, in some embodiments, the training computing system (1050) may be part of the server computing system (1030) or part of the user computing device (1010).
[0089] Meanwhile, the server computing system (1030) may include at least one processor (1031) and memory (1032). Here, the processor (1031) may be composed of at least one or a plurality of electrically connected processors among a central processing unit (CPU), a graphics processing unit (GPU), a tensor processing unit (TPU), a neural network processing unit (NPU), an application integrated circuit, an application semiconductor (ASIC), an arithmetic logic unit (ALU), a floating-point arithmetic unit (FPU), digital signal processing devices (DSPs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, and other electrical units for performing functions. For example, at least one processor (1031) may include a circuit and a transistor configured to execute instructions from memory (1032).
[0090] The memory (1032) constituting the server computing system (1030) according to the present invention may include volatile memory, non-volatile memory, fixed media, removable media, magnetic media, optical media, semiconductor media, or other types of physically durable storage media. For example, the memory (1032) may include one or more non-transient / transient computer-readable storage media such as RAM, ROM, HDD (Hard Disk Drive), SSD (Solid State Disk), SSD (Silicon Disk Drive), EEPROM, EPROM, flash memory device, magnetic disk, etc., and combinations thereof, and may include web storage of a server that performs the storage function of memory over the internet. Additionally, the server computing system (1030) may further include a data storage (data store). For example, the data storage may be composed of at least one of a relational database, a NoSQL database, a data warehouse, and a local file system.
[0091] In the memory (1032) constituting the server computing system (1030) according to the present invention, data and instructions necessary for the at least one processor (1031) to perform the operation of an application for generating high-quality data corresponding to type information for various content may be stored.
[0092] In one embodiment, the server computing system (1030) may be composed of a single device or a plurality of computing devices, and may be configured to operate according to a sequential or parallel computing architecture. Additionally, a distributed processing system may be configured with a plurality of networked devices.
[0093] Meanwhile, the training computing system (1050) may include at least one processor (1051) and memory (1052). The model trainer (1060) is a logical component that executes the training of at least one machine learning model (1020, 1040) and may be implemented in the form of hardware, firmware, or software. For example, the model trainer (1060) may be executed by the processor (1051) after loading training data (1061) stored in a storage device into memory (1052). For example, the model trainer (1060) may be configured to execute one or more operations (e.g., model training, model reconstruction, model validation, model testing) on at least one machine learning model.
[0094] The machine learning model of the present invention may include at least one of a statistical model, an algorithm, a neural network (NN), a convolutional neural network (CNN), a generative neural network (GNN), a Word2Vec model, a Bag of Words model, a TF-IDF (Document Frequency-Inverse Document Frequency) model, a GPT (Generative Pre-trained Transformer) model (or other autoregressive models), a PPO (Proximal Policy Optimization) model, a nearest neighbor model (e.g., a k-nearest neighbor model), a linear regression model, a K-means clustering model, a Q-learning model, a TD (Temporal Difference) model, a Deep Adversarial Network model, and all other types of models further described in the present invention.
[0095] Specifically, the model trainer (1060) may execute operations to train a machine learning model, and said operations may include at least one of adding, removing, and modifying model parameters. At this time, the training of the machine learning model may be at least one of supervised learning, semi-supervised learning, and unsupervised learning. In one embodiment, the training of the machine learning model may include the step of repeatedly inputting training data (1061) based on epochs and repeatedly performing the machine learning model training process configured in this way. Here, an epoch may refer to a unit in which the entire set of training data (1061) undergoes forward and backpropagation processing once. In some implementations, different levels of training methods (e.g., supervised learning, semi-supervised learning, unsupervised learning) may be used for different epochs.
[0096] The training data (1061) of the present invention may include input data and data previously output from at least one machine learning model (e.g., recursive learning feedback).
[0097] At least one parameter of a machine learning model may include at least one of a seed value, a model node, a model layer, an algorithm, a function, connections between different machine learning models, connections between parameters, machine learning model constraints, and other digital components that influence the output of the machine learning model. In this case, the model connection between different machine learning models may include or represent at least one of model parameters and relationships between models, which may be dependent or interdependent, hierarchical, and static or dynamic. The combinations and configurations of model parameters described herein may be too complex to be maintained or utilized by human cognitive abilities.
[0098] In the present invention, the machine learning parameters described according to the embodiments are not limited, and a single machine learning model may further include a plurality of model parameters.
[0099] Meanwhile, FIG. 2 illustrates an example of a block diagram of a computing device (1100) that may be included in a user computing device (1010), a server computing system (1030), and a training computing system (1050), as an embodiment of a computing system (10000) in which the present invention can be implemented.
[0100] As illustrated in FIG. 2, the computing device (1100) may include at least one application (e.g., Application 1 to Application N), and each of the at least one application may include a machine learning library and a model execution environment for performing a method to generate high-quality data corresponding to type information for various machine learning-based content. The at least one application included in the computing device (1100) may communicate with the sensor, context manager, device state manager, or additional component(s) within the computing device (1100) via an Application Programming Interface (API). In one embodiment, the at least one application may interface with device components, such as receiving sensor data or state data or transmitting prediction results to an output device via a public or private API.
[0101] Meanwhile, FIG. 3 illustrates an example of a block diagram in another aspect of a computing device (1200), which is one of the components of a computing system (10000) that performs a method for generating high-quality data corresponding to type information for various content according to an embodiment of the present invention.
[0102] A computing device (1200) according to the present invention may include at least one application (e.g., Application 1 to Application N), and at least one application may communicate with a central intelligence layer (1210). Each application may interact with a shared model within the central intelligence layer (1210) through an API (e.g., a common API).
[0103] The central intelligence layer (1210) includes one or more machine learning models and may share them among multiple applications or provide them independently to each. In one embodiment, the central intelligence layer (1210) may be integrated as part of an operating system or implemented as a separate logical layer.
[0104] Additionally, the central intelligence layer (1210) can communicate with the central device data layer (1220). The central device data layer (1220) can store data corresponding to at least one piece of content stored within the computing device (1200) and type information for said content, and can provide this as input data necessary to generate high-quality data corresponding to type information for various pieces of content. Each device component (e.g., sensor, state manager, etc.) can communicate with the central device data layer (1220) through a private API, etc.
[0105] The technology described in the present invention may be composed of a single or multiple computing devices, and a machine learning model that performs a method for generating high-quality data corresponding to type information for various content may be executed sequentially or in parallel on a single component or multiple distributed components. The data storage, machine learning model, and application may be distributed and operated locally or over a network, and these configurations can be flexibly applied to various system architectures.
[0106] The present invention aims to provide a data generation method and system capable of automatically generating high-quality data based on artificial intelligence. More specifically, the data generation system according to the present invention may be a system that generates high-quality data for improving the performance of an artificial intelligence model while encompassing various content.
[0107] For example, in the present invention, the content may include content related to at least one of a story, character design, game design, lyrics, magazine, derivative work, poem, motivational speech, podcast script, movie script, theater dialogue, rap, and TV series.
[0108] That is, the data generation system according to the present invention may be a system that generates high-quality data corresponding to (or specialized for) type information regarding various content based on artificial intelligence. Such a data generation system according to the present invention may include at least one artificial intelligence model that can be utilized depending on various situations or purposes.
[0109] For example, the artificial intelligence model may include at least one of a machine learning (ML) model, a deep learning model, a deep neural network (DNN), a language model (LM), a large language model (LLM), a super-large foundation model, a generative AI model, a transformer-based model, a supervised learning (SL) model, a reinforcement learning (RL) model, a vision-language model (VLM), and a special-purpose model (e.g., a time series forecasting model (e.g., ARIMA model, SARIMA model, etc.), a time series foundation model, a graph neural network (GNN), a multimodal model, a natural language processing (NLP) model, a computer vision model, a speech recognition or synthesis model, a recommendation system model, etc.).
[0110] The present invention aims to solve the problem of data scarcity and to automatically generate and provide data encompassing various content. Below, we will examine the data generation system according to the present invention in more detail together with the attached drawings. Figures 4 and 5 are conceptual diagrams for explaining the data generation system according to the present invention.
[0111] Meanwhile, as illustrated in FIG. 4, the data generation system (1000) according to the present invention may include at least one of an input unit (100), an output unit (200), a communication unit (300), a storage unit (400), a data generation unit (500), an evaluation unit (600), and a control unit (700). The data generation system (1000) may perform a series of data processing flows to generate new data corresponding to specific type information (e.g., second type information) that is different from type information for the input data (e.g., first type information), based on attributes related to specific type information for input data and specific content (or target content, goal content, etc.) to be generated.
[0112] However, the components of the data generation system (1000) according to the present invention are not necessarily limited thereto and may further include various hardware components that perform the same or similar roles as described in the description of the present specification.
[0113] Although not illustrated, the data generation system (1000) according to the present invention may include one or more processors, and such processors may include one or more general-purpose processors and one or more special-purpose processors (e.g., digital signal processor, tensor processing unit (TPU), graphics processing unit (GPU), neural network processing unit (NPU), application integrated circuit, application semiconductor (ASIC), field programmable gate array (FPGA), quantum processing unit (or quantum processor, QPU), etc.). One or more processors may be configured to execute instructions, computer-readable instructions, and other instructions described herein that are stored (or included) in the storage unit (400). The data generation method and system according to the present invention may perform the data processing described below in cooperation with memory and at least one processor. The processor may perform a series of operations and data processing using data and information stored in memory. In this case, memory may be a component of the storage unit (400). The above memory may be connected to a processor, and the memory may store an artificial intelligence model configured to generate data corresponding to type information for various content, and an evaluation model (or evaluator model) configured to perform quality evaluation on the generated data.
[0114] Meanwhile, the input unit (100) can be configured in various ways as a means of data input. For example, the input unit (100) can be configured to receive user input. The input unit (100) can be configured to receive user input from a user terminal (10). Here, “receiving input” may mean receiving an input signal (or selection signal) corresponding to the user’s input based on input made by the user through the input unit configuration provided in the user terminal (10).
[0115] Here, the user terminal (10) may include at least one of a mobile phone, a smartphone, a notebook computer, a laptop computer, a slate PC, a tablet PC, an ultrabook, a desktop computer, a digital broadcasting terminal, a PDA (personal digital assistants), a PMP (portable multimedia player), a navigation device, and a wearable device (e.g., a smartwatch, a smart glass, a head-mounted display).
[0116] In addition, the input unit (100) in the present invention does not necessarily mean a hardware means, but can be understood as a channel for receiving input from a user.
[0117] The input unit (100) may also be referred to as a user interface module. The input unit (100) may include a touch screen, a computer mouse, a keyboard, a keypad, a touchpad, a trackball, a joystick, a voice recognition module, or other similar devices. However, the present invention does not limit the type of input unit (100).
[0118] Here, user input may include documents, text, images (or videos), voice, etc. In this case, the data generation system (1000) may further include a module that converts voice into text.
[0119] Next, the output unit (200) can output information through an output unit configuration (e.g., a display unit, a touch screen, a speaker, etc.) provided in a user terminal (10) linked to the data generation system (1000) according to the present invention. For example, the output unit (200) can output at least one page (or service page, 2000) linked to the data generation system (1000) according to the present invention to the display unit of the user terminal (10). Additionally, the output unit (200) does not necessarily mean a hardware means, but can be understood as a channel for outputting results to a user.
[0120] Next, the communication unit (300) may be connected to a user terminal (10), a server (e.g., a central server, an external server, etc.), a device, and at least one network via a wireless or wired network, and configured to receive or transmit overall data and information necessary for the operation of the data generation system (1000) according to the present invention.
[0121] The communication unit (300) can support various communication methods according to the communication standards of the communicating device. The processor can transmit and receive data with an external server and a user terminal (10) using the communication unit.
[0122] For example, the communication unit (300) may be configured to communicate with a communication target using at least one of the following technologies: WLAN (Wireless LAN), Wi-Fi (Wireless-Fidelity), Wi-Fi (Wireless-Fidelity) Direct, DLNA (Digital Living Network Alliance), WiBro (Wireless Broadband), WiMAX (World Interoperability for Microwave Access), HSDPA (High Speed Downlink Packet Access), HSUPA (High Speed Uplink Packet Access), LTE (Long Term Evolution), LTE-A (Long Term Evolution-Advanced), 5G (5th Generation Mobile Telecommunication), Bluetooth (Bluetooth™), RFID (Radio Frequency Identification), Infrared Data Association (IrDA), UWB (Ultra-Wideband), ZigBee, NFC (Near Field Communication), Wi-Fi Direct, and Wireless USB (Wireless Universal Serial Bus).
[0123] Next, the storage unit (400, or memory) serves to store various data related to the present invention and may include one or more non-transient computer-readable storage media that can be read and accessed by at least one of one or more processors.
[0124] One or more computer-readable storage media may include volatile and non-volatile storage components, such as optical, magnetic, organic, or other memory or disk storage devices. In some examples, the storage unit (400) may be implemented using a single physical device (e.g., one optical, magnetic, organic, or other memory or disk storage device), whereas in other examples, the storage unit (400) may be implemented using two or more physical devices.
[0125] The storage unit (400) may include computer-readable instructions and additional data. The storage unit (400) may include a storage necessary to perform at least some of the methods, scenarios, and techniques described herein and at least some of the functions of the device and network.
[0126] Furthermore, at least a portion of the storage unit (400) may be a cloud storage or a cloud server. At least a portion of the data corresponding to user input received from the input unit (100) and the training data (or training data set) may be stored in the storage unit (400).
[0127] That is, the storage unit (400) is sufficient as a space where information necessary for the operation of the data generation system (1000) according to the present invention is stored, and it can be understood that there are no restrictions on the physical space.
[0128] Furthermore, the storage unit (400) can store a computer program including computer program instructions. Furthermore, the storage unit (400) can store a computer program including computer program instructions that control the operation of the system (1000) or control the operation of the control unit (700) when loaded into the processor of the system (1000).
[0129] Next, the data generation unit (500) may be configured to perform the role of generating high-quality data corresponding to type information for various content.
[0130] Specifically, the data generation unit (500) can generate high-quality data corresponding to each type information for each of the multiple different contents using an artificial intelligence model.
[0131] In one embodiment, let us assume that the query-response pair (or question-response pair) corresponding to type information for specific content is the mathematical expression illustrated in FIG. 13(a). The objective of the present invention is to expand the content types of generated data to a diverse range by generating a new query (or query) corresponding to type information for each of a plurality of contents. In the present invention, content may also be expressed as a “genre,” and type information for content may also be expressed as a “content type,” “content information,” “content type information,” “genre type,” “genre information,” or “genre type information,” etc.
[0132] The data generation unit (500) can obtain diversity of topics by sampling at least one few-shot example from a specified storage unit (e.g., storage unit (400), memory, data set (or seed data set)), combine this with content-specific attributes to reflect the unique characteristics of each content, and then generate a new query by inputting an input prompt into an artificial intelligence model.
[0133] The sampling step of the attribute subset according to the present invention, unlike simple random sampling, can be performed by referring to the distribution characteristics of each of the plurality of dimensions constituting the attribute space. For example, in the sampling step, the selection probability may be adjusted based on at least one of the frequency of occurrence, distribution balance, or inter-dimensional correlation of candidate values defined for each of the plurality of dimensions. Through such alignment of dimension distributions, the second data generated can comprehensively reflect various expression aspects of the content type of the second content without being biased toward a specific dimension or a specific attribute value.
[0134] Sampling based on the above-described dimensional distribution alignment can reduce the number of samples or processing steps required to achieve the same level of diversity compared to conventional large-scale random sampling or human curation methods for ensuring diversity of training data. Accordingly, the present invention can provide computing resource efficiency in which at least one of the number of inference calls for the artificial intelligence model, the number of evaluation calls for the evaluation model, and the storage space required for storing training data can be reduced while achieving the same distribution diversity.
[0135] Specifically, in the present invention, attributes corresponding to content can be defined as content-specific dimensions in which queries can change. For example, in the case of content related to Rap, rhythmic structure and flow may correspond to attributes (or dimensions), and in the case of content related to TV Series, elements such as narrative structure and character development may correspond to attributes (or dimensions). In the present invention, for each of a plurality of different contents (see FIG. 13 (b)), a plurality (e.g., 5 to 15) sets of attributes (see FIG. 13 (c)) can be configured (or defined). In the present invention, the attributes corresponding to each of the plurality of contents can be configured as shown in the following examples (see Table 1).
[0136] Attribute List PromptPromptThe following variables define attribute checklists for different creative artifacts. story_attributes:* Genre & Tone: Atmosphere, mood, emotional register, etc.* POV: Narrative perspective and distance, etc.* Structure: Timeline, pacing, scene flow, etc.* Characters: Count, relationships, dialogue / thought balance, etc.* Style: Action-to-description ratio, sensory detail, figurative language, etc.* poem_attributes:* Genre & Tone: Poetic category or mode (lyric, narrative, dramatic, epic, etc.); overall emotional atmosphere, tonal shifts, and consistency of mood (type or style of poetry (lyric poetry, epic poetry, dramatic poetry, epic long poetry, etc.); overall emotional atmosphere, tonal shifts, and consistency of mood).* Form & Structure: Poetic form (sonnet, free verse, haiku, etc.)); stanzaic arrangement, lineation, enjambment, repetition, and progression of imagery or ideas (poetic form (sonnet, free verse, haiku, etc.); stanzaic structure, line breaks, enjambment, repetition, development of images or ideas).* Voice & POV (Speaker & Point of View): Speaker identity and perspective; distance between poet, speaker, and reader; use of persona or multiple voices; intimacy and emotional stance (speaker identity and perspective; distance between poet, speaker, and reader; use of persona or multiple speakers; intimacy and emotional stance).* Sound & Rhythm (Sound & Rhythm): Meter, cadence, rhyme scheme, alliteration, assonance, consonance, and overall musicality; how sound patterns reinforce meaning or emotion (phonological elements such as rhythm, meter, prosodic structure, alliteration, rhyme, and consonant repetition, and overall musicality; how these sound patterns reinforce meaning or emotion).* Imagery & Symbolism: Sensory detail, figurative language (metaphor, simile, personification, etc.), recurring symbols, and the way images build emotional or thematic resonance.* Theme & Meaning: Central ideas, philosophical or emotional core, subtext, moral or existential questions; evolution of meaning through the poem's progression.* Language & Diction: Word choice, register, syntax, concision or excess, ambiguity, and texture of language; Interplay between denotation and connotation (word choice, stylistic level, sentence structure, conciseness or excess, ambiguity, linguistic texture, interaction between denotative and connotative meaning).* Pacing & Movement: Flow between lines and stanzas, pauses and breaks (caesura, enjambment), rhythmic tension and release, and narrative or emotional trajectory.* Emotion & Affect: The emotional journey for both speaker and reader; how tone, imagery, and rhythm evoke and modulate feeling.* Style & Voice: Distinctive stylistic fingerprint—clarity vs. abstraction, irony vs.sincerity, density of imagery, and authorial or era-specific voice (unique stylistic characteristics—clarity vs. abstraction, irony vs. authenticity, image density, authorial or era-specific voice).lyrics_attributes:* Genre & Tone: Musical style, emotional atmosphere, mood shifts, sincerity vs. irony, playfulness vs. gravity, and tonal consistency or contrast (musical style, emotional mood, mood shifts, authenticity versus irony, playfulness versus seriousness, tonal consistency or contrast).* Structure & Form: Verse-chorus arrangement, bridge placement, pre-chorus usage, repetition patterns, stanza length, rhyme scheme (ABAB, AABB, etc.), and structural variations or breaks.* Voice & Perspective: First-person, second-person, third-person, or shifting POV; direct address vs. observational stance; persona or character adopted by the speaker.* Imagery & Figurative Language: Use of metaphor, simile, symbolism, personification; sensory details (visual, auditory, tactile); concrete vs. abstract language; and vividness of imagery.* Rhythm & Sonic Quality: Syllable count, meter, cadence, internal rhyme, alliteration, assonance, consonance; how lyrics sync with melody; Breathability and singability (number of syllables, meter, internal rhyme, alliteration, rhyme, consonant repetition; how the lyrics harmonize with the melody; ease of pronunciation and singability).* Theme & Message: Central ideas, emotional core, social commentary, personal reflection, storytelling intent, and underlying meaning or subtext.* Narrative & Storytelling: Presence of plot or characters, chronological vs. fragmented storytelling, scene-setting, and narrative arc or resolution.* Wordplay & Craft: Clever phrasing, double meanings, puns, cultural references, vocabulary range (colloquial vs. poetic), and lyrical complexity or simplicity. * Repetition & Hook: Use of refrains, memorable lines, earworm potential, and how repetition reinforces theme or emotion. * Emotional Arc & Dynamics: Progression of feeling throughout the song, tension and release, vulnerability vs. strength, and emotional payoff(emotional changes throughout the song, tension and release, contrast between vulnerability and strength, emotional climax).movie_script_attributes(movie script attributes):* Genre & Tone(Genre & Tone): Emotional atmosphere, thematic flavor, overall mood, and tonal consistency throughout the script(emotional atmosphere, thematic flavor, overall mood, and tonal consistency throughout the script).* Structure & Pacing: Act divisions (3-act or alternative), plot points, rhythm of tension and release, scene sequencing, narrative momentum, and buildup / payoff patterns.* Characterization & Dialogue: Character arcs, relationships, motivations, subtext; balance between spoken lines, silence, and action-driven characterization; distinctive voice and rhythm (character arc, relationships, motivation, hidden meaning; balance between dialogue, silence, and action-oriented expression; unique tone and rhythm).* Visual Storytelling & Cinematic Flow: Use of cinematic imagery, blocking, stage direction, description density, visual readability, shot rhythm potential, and how the script translates to screen.* Conflict & Stakes: Core dramatic question, layers of internal and external conflict, and how stakes evolve or escalate.* Theme & Message: Underlying ideas, motifs, or moral questions that drive the story's deeper meaning.* POV & Focus: Point of view (omniscient, limited, alternating), emotional distance from characters, and audience alignment.* Scene Dynamics: Scene purpose, transitions, interplay between exposition, tension, and release.* Style & Voice: Distinctive narrative voice, use of humor or symbolism, and overall screenwriting flair.* rap_attributes:* Genre & Tone: Sub-style (boom bap, trap, drill, conscious, battle, experimental); emotional atmosphere (aggressive, reflective, playful, defiant); sincerity vs.irony; balance between swagger, humor, and intensity; tonal evolution through the track (subgenres (boom bap, trap, drill, conscious, battle, experimental, etc.); emotional atmosphere (aggressive, reflective, playful, challenging); authenticity versus irony; balance of swagger, humor, and intensity; tonal changes throughout the track).* Flow & Delivery: Rhythm, cadence, and syllabic density; syncopation and off-beat phrasing; use of pauses and breath control; vocal tone, emphasis, dynamics, and personality; charisma, confidence, and emotional projection in performance (rhythm, beat, syllabic density; syncopation and off-beat phrasing; use of pauses and breath control; vocal tone, emphasis, dynamics, and personality; charisma, confidence, and emotional projection in performance).* Structure & Form: Arrangement of intro, verses, hooks, bridges, outros; use of breakdowns or beat switches; bar count per section (16-bar verse, 8-bar hook); patterns of repetition; freestyle vs. written structure (composition of intro, verse, hook, bridge, outro; use of breakdowns or beat transitions; number of bars per section (e.g., 16-bar verse, 8-bar hook); repetition patterns; freestyle vs. written structure).* Rhyme & Wordplay: Complexity of rhyme schemes (multisyllabic, slant, internal, compound); punchlines, puns, double entendres; alliteration, assonance, consonance; creative phrasing and manipulation of language.* Lyrics & Content: Central themes (identity, struggle, power, success, politics, humor); narrative presence or concept; storytelling vs. abstract expression; emotional depth vs. bravado; Authenticity of voice (central theme (identity, struggle, power, success, politics, humor); presence of narrative or concept; story-centered vs. abstract expression; emotional depth vs. bravado; authenticity of the voice).* Flow-Beat Relationship: How the rapper interacts with the instrumental (pocketing, anticipation, syncopation); contrast or complement between vocal rhythm and percussion; adaptability to beat changes or tempo shifts.* Persona & Voice: Character, attitude, and stance adopted; first-person vs. collective or third-person perspective; consistency of persona; rhetorical stance (preacher, philosopher, rebel, clown); emotional transparency vs. performative mask. * Sound & Sonic Texture: Timbre of voice; layering, ad-libs, harmonics, background vocals; mixing of vocal tracks; production texture (sample-based, synthetic, lo-fi, polished); interplay of vocal tone and instrumental mood.* Cultural & Contextual Layer: References to social, political, and cultural realities; authenticity and locality (slang, dialect, region); intertextual nods to hip-hop history; commentary on fame, race, gender, and power; placement within hip-hop tradition or innovation. * Hook & Memorability: Construction of refrains or catchphrases; melodic or rhythmic hooks; repetition strategy; call-and-response potential; how the hook reinforces theme or emotion; overall memorability and live-readiness.* Emotional & Narrative Arc: Progression of tension, tone, and confidence across verses; climaxes and resolutions; emotional pay-off or transformation; contrast between beginning and ending states; overall energy arc across the track. * Character Design Attributes: * Physical Design & Appearance: Visual presentation, distinctive features, costume / clothing choices, color palette, silhouette recognition, age indicators, body language, and how appearance reflects personality or role. * Personality & Psychology: Core traits, temperament, values, beliefs, fears, desires, contradictions, emotional range, consistency vs. complexity, and psychological depth or archetypal simplicity (core disposition, temperament, values, beliefs, fears, desires, contradictions, emotional range, consistency vs. complexity, psychological depth or archetypal simplicity).* Backstory & History: Past experiences, origin, formative events, trauma or triumph, family background, cultural context, and how history shapes present behavior and motivations. * Goals & Motivations: Driving desires, immediate objectives, long-term aspirations, internal vs. external goals, conflicting motivations, and what the character wants vs. needs. * Relationships & Dynamics: Connections to other characters, relationship arcs, power dynamics, alliances and rivalries, romantic interests, mentors / students, and social positioning within the narrative world.* Character Arc & Development: Growth trajectory, transformation or stasis, challenges faced, lessons learned, regression or progression, turning points, and whether the arc is redemptive, tragic, or circular. * Voice & Dialogue Style: Speech patterns, vocabulary, accent or dialect, verbal tics, formality level, humor style, and how they express themselves differently in various situations. * Skills & Abilities: Competencies, talents, weaknesses, learning curve, expertise areas, combat or magic systems (if applicable), and how abilities tie to identity or plot function.* Role & Function: Protagonist, antagonist, mentor, comic relief, foil, catalyst; narrative purpose; whether they drive the plot or serve thematic functions; archetypal role.* Moral Alignment & Ethics: Value system, ethical boundaries, willingness to compromise, gray areas, heroic vs. villainous qualities, and moral complexity or simplicity.* Medium-Specific Execution: For novels—internal monologue depth, description balance; for film / TV—actor embodiment, visual storytelling; for games—player agency, customization options, dialogue trees; for comics / manga—panel presence, visual expressiveness(novels—depth of internal monologue, descriptive ratio; film / TV—actor's embodiment, visual expression; games—player choice, customization options, dialogue tree; comics / manga—panel presence, visual expressiveness).* Cultural & Symbolic Dimensions: Cultural representation, stereotypes avoided or invoked, symbolic meaning, thematic embodiment, and what the character represents beyond themselves. * Memorability & Iconicity: Distinctive hooks, quotability, visual or behavioral signatures, fan engagement potential, and lasting impression factors. * Consistency & Believability: Internal logic, behavioral consistency, justified exceptions, realistic responses to situations, and balance between predictability and surprise. * Game_design_attributes: * Core Mechanics & Systems: Rule sets, core gameplay loops, action-consequence relationships, resource management, progression systems, difficulty scaling, and mechanical depth vs.Accessibility balance (rule system, core gameplay loop, action-result relationships, resource management, progression system, difficulty adjustment, balance between mechanical depth and accessibility).* Player Agency & Choice: Meaningful decisionmaking opportunities, branching paths, consequence visibility, player freedom of expression, constraint design, and the balance between guidance and autonomy.* Challenge & Difficulty: Learning curve design, skill ceiling and floor, failure states and retry mechanics, difficulty modes or adaptive systems, challenge pacing, and risk-reward balance.* Feedback & Communication: UI / UX clarity, audio-visual feedback systems, tutorial and onboarding design, information hierarchy, player goal communication, and system transparency vs. mystery(UI / UX clarity, audio-visual feedback system, tutorial and onboarding design, information structure, player goal delivery, system transparency vs. mystery).* Narrative Integration: Story delivery methods (environmental, cutscenes, emergent), ludonarra-tive harmony or dissonance, player role in narrative, world-building techniques, and narrative pacing vs. gameplay pacing. harmony or dissonance, player role in narrative, world-building techniques, and narrative pacing vs. gameplay pacing. harmony or dissonance, player role in narrative, world-building techniques, and narrative pacing vs. Gameplay pacing (story delivery method (environmental, cutscene, emergent), ludonartic harmony or inconsistency, player role within the narrative, world-building techniques, balance between narrative pacing and gameplay pacing).* Spatial & Level Design: Environment layout, navigational flow, landmark and waypoint design, spatial storytelling, exploration incentives, and architectural purpose (combat arenas, puzzle spaces, narrative zones).* Aesthetic & Atmosphere: Art direction, color palette, sound design, musical scoring, visual clarity vs.artistic expression, thematic cohesion, and emotional tone setting (art direction, color palette, sound design, music, balance of visual clarity and artistic expression, thematic cohesion, formation of emotional atmosphere).* Player Motivation & Engagement: Intrinsic vs. extrinsic reward structures, goal-setting mechanisms, achievement systems, discovery and surprise elements, replay value design, and player investment hooks (intrinsic / extrinsic reward structures, goal-setting mechanisms, achievement systems, discovery and surprise elements, replay value design, player investment elements).* Social & Multiplayer Dynamics (if applicable): Cooperation vs. competition balance, communication tools, social interaction design, community features, matchmaking systems, and player behavior incentives (balance between cooperation and competition, communication tools, social interaction design, community features, matchmaking systems, player behavior incentives).* Accessibility & Inclusivity: Control customization, difficulty accommodations, sensory alternatives (subtitles, colorblind modes), input flexibility, and barrier reduction for diverse player needs. * Economy & Progression: Experience and leveling systems, currency and trade mechanics, unlock structures, power curve design, grind vs. reward balance, and monetization integration (if applicable). * Innovation & Convention: Genre adherence vs. subversion, mechanical originality, quality-of-life features, industry standard expectations, and creative risk-taking vs. familiarity(Genre convention compliance vs. subversion, originality of mechanism, convenience features, industry standard expectations, creative experimentation vs. familiarity).tv_series_attributes(TV Series Attributes):* Genre & Tone(Genre & Tone): Type of series (drama, comedy, thriller, sci-fi, etc.); overall mood and emotional register; tonal consistency or shifts across episodes or seasons; balance of realism vs. stylization; interplay of humor, tension, and pathos (Series type (drama, comedy, thriller, sci-fi, etc.); overall mood and range of emotional expression; tonal consistency or shifts between episodes and seasons; balance of realism versus stylization; interplay of humor, tension, and pathos).* Narrative Structure & Pacing: Season and episode structure; serialized vs. episodic storytelling; pacing within and across episodes; use of cliffhangers, flashbacks, or parallel timelines; rhythm of plot progression and emotional arcs (Season and episode structure; distinction between serial and episodic narratives; pace of development within and between episodes; use of cliffhangers, flashbacks, and parallel timelines; rhythm of plot progression and emotional flow).* World-Building & Setting: Level of immersion and detail in the setting; temporal and geographic scope; rules and logic of the fictional world; sociocultural and political context; Integration of environment into character and plot development (level of background immersion and detail; temporal and geographical scope; rules and logic of the fictional world; sociocultural and political context; influence of the environment on character and plot development).* Characters & Relationships: Range and depth of main and supporting characters; development arcs; relational dynamics and conflicts; dialogue realism and subtext; presence of ensemble balance or central protagonist focus; representation and diversity.* Theme & Subtext: Central ideas and recurring motifs; social, moral, or philosophical commentary; subtextual tensions (e.g., identity, power, morality); symbolic or allegorical layers; evolution of thematic material over time.* Cinematography & Visual Style: Camera work, framing, lighting, color palette, and visual motifs; consistency of visual language; symbolic use of composition; aesthetic tone (gritty, polished, sur-real, documentary, etc.); how visuals support emotion and narrative (camera direction, composition, lighting, color palette, visual motifs; consistency of visual language; symbolic expression through composition; aesthetic style (rough and realistic, sophisticated, surreal, documentary style, etc.); how visual elements enhance emotion and narrative).* Editing & Narrative Rhythm: Scene transitions and temporal flow; cross-cutting, montage, and juxtaposition techniques; pacing of visual storytelling; continuity vs. fragmentation; use of rhythm to enhance tone or tension.* Sound & Music Design: Score and soundtrack function; diegetic vs. non-diegetic sound; use of silence and ambient sound; audio motifs tied to characters or themes; Alignment of musical tone with emotional or narrative beats (role of background music and soundtrack; digetic / non-digetic sound; use of silence and ambient sounds; acoustic motifs connected to characters or themes; alignment of musical tone with emotion or narrative flow).* Dialogue & Writing Style: Voice authenticity; linguistic tone and rhythm; use of wit, subtext, or exposition; writing economy and balance between show-and-tell; distinctive speech patterns for characters; stylistic trademarks of the writers.* Direction & Performance: Acting quality and consistency; chemistry among cast; directorial interpretation and tone management; visual and emotional coherence across directors or showrunners; performance nuance and authenticity.* Production Design & Costume: Set, props, wardrobe, and makeup as storytelling tools; historical or genre accuracy; visual symbolism in design choices; how aesthetic elements reflect character psychology or thematic resonance (narrative function of sets, props, costumes, and makeup; historical and genre accuracy; visual symbolism in design choices; how aesthetic elements reflect character psychology or thematic meaning).* Audience Engagement & Reception: Emotional impact, rewatchability, and fandom dynamics; critical vs. popular response; cultural relevance and influence; use of cliffhangers or mysteries to sustain attention; viewer identification and empathy. * Cultural & Contextual Relevance: Reflection of contemporary social issues; representation politics; commentary on cultural narratives or media tropes; evolution of audience expectations; intertextual references and meta-commentary. * Series Cohesion & Evolution: Narrative and stylistic consistency across seasons; adaptation to shifts in tone, cast, or direction; resolution quality; balance between innovation and continuity; long-term payoff of arcs and themes (consistency of narrative and style across seasons; adaptation to changes in tone, cast, and direction; completeness of the ending; balance between innovation and continuity; long-term completeness of narrative arcs and themes).derivative_work_attributes:* Source Relationship: Degree of transformation or reinterpretation of the original work; faithfulness vs. innovation; how much the derivative depends on source context.* Medium & Form Shift: Changes in artistic medium (e.g., novel to film, song to visual art) and how form alters meaning, pacing, or sensory experience.* Thematic Continuity & Divergence: Retained, expanded, or subverted themes; how core ideas evolve or take on new significance (themes maintained, expanded, or subverted; how core ideas change or acquire new meaning).* Tone & Mood Adaptation: Consistency or contrast in emotional tone between original and derivative; how mood is recontextualized or intensified (consistency or contrast in emotional tone between the original and derivative; how the mood is recontextualized or intensified).* Character or Motif Transformation: Reimagining of characters, symbols, or recurring motifs; shifts in perspective, motivation, or narrative function. * Structural Adaptation: How narrative or compositional structure changes; compression, expansion, or rearrangement of elements. * Audience & Intent: New target audience, inter-pretive stance, or creator's purpose in adaptation; homage, critique, parody, or expansion. * Intertextuality & Reference: Explicit or implicit references to the source; use of quotation, allusion, or meta-commentary (explicit or implicit reference to the original work; use of quotations, allusions, or meta-commentary).* Cultural & Temporal Reframing: How shifts in culture, time, or setting alter meaning; modernization, localization, or historical reinterpretation (the impact of changes in culture, time, or setting on meaning; modernization, localization, or historical reinterpretation).* Aesthetic & Stylistic Translation: Preservation or transformation of artistic style, tone, rhythm, or imagery across media. * Originality & Authorship: Balance between borrowed and original material; creative agency and ethical boundaries of adaptation. * Emotional Resonance & Impact: Comparative emotional effect; how the derivative amplifies, contrasts, or complicates the original's emotional core.podcast_attributes:* Format & Structure: Episode type (interview, solo, co-hosted, narrative, roundtable), segment organization, intro / outro placement, ad break positioning, and overall flow.* Host Voice & Persona: Tone (conversational, authoritative, intimate, humorous), personality projection, relatability, authenticity, and consistency of character. * Conversational Dynamics: Turn-taking patterns, interruption style, back-and-forth rhythm, active listening cues, and natural vs. scripted feel (speech sequence patterns, interruption style, rhythm of exchange, active listening cues, naturalness vs. script-based progression).* Pacing & Timing: Episode length considerations, beat distribution, momentum building, strategic pauses, topic transition speed, and energy fluctuations.* Audio-Specific Elements: Stage directions for sound effects, music cues, vocal emphasis indicators, breath marks, laugh lines, and production notes.* Content Depth & Research: Fact density, source citation approach, expertise demonstration, preparation level, and balance of depth vs. Accessibility (information density, source citation style, expression of expertise, level of preparation, balance of depth and accessibility).* Audience Engagement: Direct address frequency, question-posing, call-and-response elements, community references, listener interaction prompts, and inclusive language.* Storytelling Techniques: Anecdote placement, narrative hooks, cliffhangers, scene-setting, character development (for narrative podcasts), and dramatic tension.* Language & Register: Formality level, jargon usage, code-switching, colloquialisms, verbal tics, filler words (strategic or natural), and vocabulary accessibility.* Topic Development: Argument structure, idea progression, tangent management, callback usage, thematic threading, and point-of-view clarity.* Humor & Personality: Joke placement, comedic timing notes, banter style, wit vs. slapstick, self-deprecation, and tonal lightness vs. Seriousness (placement of jokes, comedy timing, conversation style, wit versus slapstick, self-deprecation, lightness and seriousness of tone).* Educational Intent: Explanation clarity, analogy usage, complexity scaffolding, recap frequency, and knowledge assumption about the audience.* Emotional Resonance: Vulnerability moments, empathy building, emotional authenticity, tonal warmth, and connection-building strategies.* Technical Craft: Readability for voice, breath phrase length, tongue-twister avoidance, phonetic clarity, and ease of delivery.* Structural Signposting: Episode roadmapping, segment announcements, time markers, topic previews, and navigation aids for listeners.theater_dialogue_attributes:* Genre & Tone: Overall dramatic style (tragedy, comedy, realism, absurdism, etc.), mood, emotional intensity, and tonal consistency or contrast throughout the dialogue.* Structure & Pacing: Scene organization, rhythm of exchanges, use of pauses and silences, escalation and release of tension, and progression of dramatic beats.* Characterization & Relationships & Relationships): How dialogue reveals character traits, motivations, power dynamics, emotional undercurrents, and interpersonal conflicts or alliances.* Voice & Language Style: Distinctiveness of each character's speech patterns, diction, register (formal vs. colloquial), idiomatic expressions, and subtext through tone or delivery. * Conflict & Tension: Sources of dramatic conflict (internal or external), how tension is built and sustained through dialogue, and resolution or lack thereof. * Subtext & Implication: What is unsaid or implied beneath the surface of the dialogue, use of irony, double meanings, and layered communication between characters. * Rhythm & Timing: Verbal pacing, use of interruptions, overlaps, comedic timing. and how rhythm supports emotional or thematic beats (speech rate, interruptions, dialogue overlap, comedic timing, how rhythm supports emotional or thematic development).* Theme & Message: Underlying ideas or philosophical questions explored through dialogue, moral dilemmas, and reflections on society, identity, or human nature. * Staging & Performative Potential: How dialogue interacts with physical movement, gestures, stage directions, and spatial dynamics to enhance meaning and emotion. * Emotional Arc & Transformation: Development of emotional states within the dialogue, turning points, revelations, and shifts in power or understanding between characters. * Speech_attributes: * Purpose & Intent: Primary goal (inspire, challenge, comfort, activate), target audience, desired behavioral or emotional outcome, and specificity of call-to-action(main objective (inspiration, challenge, comfort, call to action), target audience, expected behavior or emotional response, specificity of the call to action).* Tone & Emotional Register: Energy level (highintensity vs. calm reassurance), sincerity vs. performance, warmth vs. authority, urgency vs. Patience, and tonal consistency or strategic shifts (energy level (high intensity vs. calm stability), authenticity vs. performativity, warmth vs. authority, urgency vs. patience, tonal consistency or strategic shifts).* Structure & Organization: Opening hook strength, logical flow of ideas, use of setup-payoff patterns, climactic moments, transitional phrases, and conclusion effectiveness.* Rhetorical Devices: Use of repetition (anaphora, epistrophe), parallelism, rhetorical questions, rule of three, antithesis, metaphor, and other persuasive techniques.* Voice & Perspective: First-person credibility building, second-person direct address, inclusive 'we' us-age, personal story vs.Universal truth balance, and speaker persona (coach, mentor, peer, authority) (Building trust through first person, second-person direct appeal, use of 'we', balance between personal stories and universal truth, speaker persona (coach, mentor, peer, authority)).* Evidence & Support: Use of personal anecdotes, statistics, quotes from others, historical examples, hypothetical scenarios, and balance between emotional and logical appeals.* Language & Diction: Vocabulary complexity, conversational vs. formal register, use of power words, concrete vs. abstract language, and cultural or contextual references (vocabulary difficulty, spoken / written level, use of emphasized words, concrete / abstract expressions, cultural and contextual references).* Pacing & Rhythm: Sentence length variation, strategic pauses, momentum building, speed of delivery indicators, and breathability of language.* Emotional Arc: Emotional journey design, pain point acknowledgment, hope injection timing, tension and release patterns, and crescendo placement. * Authenticity & Credibility: Vulnerability vs. strength balance, relatability markers, expertise demonstration, admission of struggle or failure, and trust-building elements. * Call-to-Action: Clarity and specificity, timing within speech, actionability level, inspirational vs. practical balance, and empowerment messaging (clarity and specificity, positioning within the utterance, feasibility, balance of inspiration and practicality, empowerment messaging).* Imagery & Visualization: Use of vivid descriptions, future-pacing techniques, sensory language, metaphorical journeys, and mental picture creation.magazine_attributes:* Publication Type & Format: Frequency (weekly, monthly, quarterly), print vs.Digital, size / dimensions, page count range, binding style, and overall physical presentation or digital interface (publication frequency (weekly, monthly, quarterly), print vs. digital, size / specifications, page count range, binding method, overall physical composition or digital interface).* Editorial Focus & Niche: Subject matter specialization (news, lifestyle, trade, academic, etc.), target demographic, geographic scope, and editorial mission or positioning.* Visual Design & Layout: Typography choices, grid systems, white space usage, image-to-text ratio, color palette, cover design philosophy, and consistency of visual identity across issues.* Content Structure & Organization: Section breakdown (departments, features, columns), front-of-book vs. back-of-book content, table of contents design, navigation aids, and content hierarchy (section organization (corners, features, columns), distinction between front and back content, table of contents design, navigation elements, content hierarchy).* Article Types & Mix: Balance of long-form features, short news items, profiles, reviews, opinion pieces, service journalism, and recurring columns; editorial calendar patterns. * Voice & Editorial Tone: Authoritative vs. conversational, formal vs. casual, objective vs. subjective, aspirational vs. accessible, and consistency across different sections. * Photography & Illustration: Style of imagery (documentary, stylized, conceptual), commissioned vs. stock, illustration usage, photo editing approach, and visual storytelling techniques.* Advertising Integration: Ad-to-editorial ratio, advertorial presence, ad placement strategy, visual distinction between ads and content, and sponsor relationships. * Audience Engagement: Letters to the editor, reader submissions, interactive elements, social media integration, community-building features, and reader service components. * Writing Style & Quality: Sentence structure complexity, vocabulary level, reporting depth, fact-checking rigor, use of sources, and overall journalistic or creative standards. * Thematic Consistency: Recurring topics or angles, seasonal patterns, ideological perspective or bias, cultural values reflected, and editorial point of view(recurring themes or perspectives, seasonal patterns, ideological viewpoints or biases, reflected cultural values, editorial direction).* Masthead & Contributors: Editorial team structure, staff vs. freelance writer ratio, contributor diversity, editor's letter presence, and editorial leadership style.
[0137] This Table 1 can also be understood as a prompt containing a list of attributes by content.
[0138] Subsequently, when the data generation unit (500) is given a sampled few-shot example set (see (d) in FIG. 13) and an attribute subset (see (e) in FIG. 13), it can construct a structured prompt (or structured input data, structured input information, etc.) through a template (see (f) in FIG. 13) and input the prompt into an artificial intelligence model to generate a new query. This can be represented as shown in (g) in FIG. 13.
[0139] To explain the data generation process according to the present invention in more detail, first, the data generation unit (500) can perform at least one filtering to ensure the quality of the data. Here, performing filtering can also be understood as creating a refined specific data set by removing at least one of harmful content and irrelevant content (or unnecessary content) included in a specific data set (e.g., seed data set).
[0140] For example, the data generation unit (500) can filter data included in a specific data set (e.g., seed data set). The data generation unit (500) can remove harmful content containing harmful or inappropriate content among the data included in the specific data set through safety filtering. Additionally, the data generation unit (500) can perform irrelevant content removal among the data included in the specific data set (e.g., seed data set) that is not related to the topic. For example, irrelevant content may include meta comments or subreddit announcement text.
[0141] In this case, the data generation unit (500) can remove harmful content and irrelevant content included in a specific data set using a prompt. The prompt used for removing such harmful content and irrelevant content can be configured as in the following example (see Table 2).
[0142] Safety and Irrelevant Content Analysis PromptTask Analyze the given topic and body for safety issues and irrelevant content. Safety Check Flag content as unsafe(is_safe: False) ONLY if it contains explicit: * Hate speech or harassment * Graphic violence or harm instructions * Sexual exploitation content * Self-harm / suicide encouragement * Dangerous illegal activity instructions Be lenient—only flag truly dangerous content.Irrelevant Content Check: Identify and remove only content completely unrelated to the main writing, such as: * Author self-promotions * External links to personal sites / subreddits * Signature blocks or footnotes promoting other work * Separators (like “” followed by promotional text) Preserve all story content, dialogue, and narrative elements intact.Output Format {"is_safe": true / false,"has_irrelevant_content": true / false,"refined_topic": "cleaned topic if safe but has irrelevant content, else empty string","refined_body": "cleaned body if safe but has irrelevant content, else empty string"}Rules * If content is safe AND has no irrelevant content: both refined fields = "" * If unsafe: refined fields = "" (content should not be processed) * If safe BUT has irrelevant content: provide cleaned versions in refined fields * CRITICAL: Keep original content exactly as-is. Only DELETE irrelevant parts Do not add, edit, rephrase, or modify anything else(Important: The original content must be preserved exactly as is. Delete only irrelevant parts; do not add, edit, reorganize, or modify).Input(Input)# Topic{topic}# Body{body}.
[0143] That is, the data generation unit (500) inputs a prompt such as the one in Table 2 above into an artificial intelligence model to perform safety filtering and removal of irrelevant content for a specific data set, and thereby obtains a refined specific data set.
[0144] The above prompt may be configured as structured input information by combining the above first data, the above type information, and the above attribute data according to predetermined serialization rules. The serialization rules may define at least one of the position of each information element within the input sequence, the type of delimiter or token separating the information elements, and a notation format according to a format recognizable by the artificial intelligence model. By configuring such structured input information, the artificial intelligence model can distinguish and process which part within the input sequence corresponds to an example, which part corresponds to a content type indication of the target to be generated, and which part corresponds to structural and stylistic constraints.
[0145] In one embodiment of the present invention, the structured input information may include a few-shot example portion transmitted to the artificial intelligence model through a first information path and an attribute data portion transmitted to the artificial intelligence model through a second information path, and the artificial intelligence model may obtain a learning signal of a representation style from the input through the first information path and simultaneously obtain structural constraints of the content to be generated from the input through the second information path to output the second data. This integrated structure within a single input sequence of dual information paths can reduce at least one of the number of model calls and the amount of intermediate data storage required compared to the conventional method of processing representation styles and structural constraints separately in a separate learning step or a separate model.
[0146] The above data set refers to a set of data including the second data, and in one embodiment, the data set may be configured to include a plurality of second data corresponding to the type information.
[0147] And, for each template instance, the data generation unit (500) may randomly sample at least some query-response pairs from a specific refined dataset and use them as few-shot examples within the prompt template (see S501 and S503 in FIG. 5). This process can also be understood as a few-shot sampling process (e.g., “Few-Shot Sampling”). For example, the sampled few-shot examples provide the artificial intelligence model with various topics and writing styles, thereby promoting thematic diversity of the generated queries.
[0148] Subsequently, the data generation unit (500) may randomly sample attributes for each content (e.g., “Genre Attribute Sampling”, see S505 in FIG. 5). The data generation unit (500) may sample (or select) at least some of the attributes corresponding to the target content (or target content). These attributes may define (or reflect) structural features or stylistic requirements (e.g., rhythm, narrative structure, dialogue format, etc.) of the target content.
[0149] In one embodiment, the data generation unit (500) samples at least one value within the range of the total number of attributes defined for the content, and can randomly select a number of attributes corresponding to the sampled value from the set of attributes of the genre. That is, a subset of attributes satisfying the condition that the number of attributes becomes k (see (h) in FIG. 13) can be randomly selected. When the number of attributes (see (i) in FIG. 13) is 0, the query remains in an open state with minimal constraints, and as the number of attributes increases, the query may gradually become more constrained by adding explicit requirements.
[0150] In another embodiment, the data generation unit (500) is not limited to randomly selecting the attribute subset, but may sample the attribute subset so as to be aligned with the attribute distribution characteristics corresponding to the target content. The attribute distribution characteristics may be defined based on at least one of the frequency of occurrence, importance weight, or mutual combination frequency of attributes defined for the content type. For example, the data generation unit (500) may sample the attribute subset corresponding to the number of samples within the total number range of attributes defined for the content in at least one of weighted sampling, stratified sampling, or distribution-aligned sampling so as to be aligned with the attribute distribution characteristics. Through such distribution-aligned sampling, the attribute data generated can consistently express the structural and stylistic characteristics of the target content while reflecting the statistical characteristics of the content type.
[0151] Furthermore, the data generation unit (500) can generate an attribute-based query from the prompt (e.g., “Attribute-Guided Query Generation”, see S507 in FIG. 5). The data generation unit (500) can configure a prompt including at least one of the sampled few-shot example, type information for the target content, and the sampled attribute, and input this into an artificial intelligence model to generate a query corresponding to the target content. In this case, multiple queries (e.g., 5) can be generated for each prompt.
[0152] The above query can be generated to have various themes and expressions while maintaining the characteristics of the target content. That is, the above query can be generated in a form that reflects the structure and style of the target content while maintaining the thematic diversity of the input content.
[0153] Finally, through the process described above, the data generation unit (500) can generate a query corresponding to each of the multiple different contents (or type information for each of the multiple contents) (e.g., “Generated Multi-Genre Creative Writing Queries”, see S509 in FIG. 5).
[0154] Meanwhile, the evaluation unit (600) may be configured to perform a quality evaluation of the generated data. The generated data may include query-response pairs in which a query and a response to the query form a pair.
[0155] For example, the evaluation unit (600) can perform a quality evaluation of query-response pairs using an evaluation model to filter out low-quality data, and select at least one query-response pair that satisfies a preset standard based on the quality evaluation.
[0156] Furthermore, the selection criteria based on the quality evaluation described above are not limited to specific statistical values and can be flexibly set according to various criteria. For example, the selection criteria may include statistical criteria defined based on the distribution of evaluation scores for all query-response pairs, and these statistical criteria may be set to include at least one of the deviation from the mean, percentile, quartile, variance, or standard deviation-based range. Additionally, the selection criteria may be dynamically adjustable rather than fixed values and may be set differently depending on the characteristics of the data set, the type of content, or the purpose of application. Furthermore, the selection criteria may be set by combining multiple evaluation criteria rather than a single criterion, and may include, for example, minimum score criteria, relative ranking criteria, or weighted sum score criteria of multiple evaluation items.
[0157] Additionally, the above-mentioned preset criteria may include a quality threshold that is compared with the quality score calculated by the evaluation model. In one embodiment, the data generation system (1000) may select query-response pairs or data in which the quality score calculated from the evaluation model is greater than or equal to the quality threshold, and exclude query-response pairs or data in which the quality score is less than the quality threshold. The quality threshold is an absolute score criterion (e.g., 1
[0158] In addition, the above evaluation criteria may be set to include at least one of a predefined rule-based method, a machine learning-based method, or a user-defined criterion.
[0159] On the other hand, query-response pairs that do not satisfy the pre-established criteria may be excluded (i.e., query-response pairs that receive a score 2 standard deviations or less below the mean are excluded).
[0160] In this regard, the data set (900, or final data set) generated (or constructed) in the present invention may include query-response pairs selected based on the quality evaluation.
[0161] Next, the control unit (700) can perform the role of controlling the overall operation of the data generation system (1000) related to the present invention. The control unit (700) can process signals, data, information, etc. that are input or output through the components of the data generation system (1000) described above, or perform a series of data processing to provide or process appropriate information and functions to the user. The control unit (700) can be physically implemented by the processor described above.
[0162] Meanwhile, as described above, the present invention can provide a data generation system (1000) that generates high-quality data for improving the performance of an artificial intelligence model while encompassing various content. The data generation system (1000) according to the present invention can generate high-quality data corresponding to (or specialized for) each type information for each of a plurality of different content by using an artificial intelligence model.
[0163] Hereinafter, together with the attached drawings, we will examine in more detail the data generation method performed by the data generation system (1000) according to the present invention. FIG. 6 is a flowchart for explaining the data generation method according to the present invention, and FIGS. 7, 8, 9, 10, and 11 below, together with the attached drawings, we will examine in more detail the data generation method performed by the data generation system (1000) according to the present invention. FIG. 6 is a flowchart for explaining the data generation method according to the present invention, and FIGS. 7, 8, 9, and 12 are conceptual diagrams for explaining the data generation method according to the present invention. Furthermore, FIGS. 13 and 14 are drawings illustrating mathematical expressions for explaining the data generation method according to the present invention.
[0164] Meanwhile, as illustrated in FIG. 6, the data generation method according to the present invention may include the step of receiving at least one first data corresponding to a first content (S610), the step of receiving type information for a second content different from the first content and generating at least one attribute data related to the second content based on the type information (S620), and the step of generating at least one second data corresponding to the type information based on an artificial intelligence model using at least one of the first data and the attribute data (S630). Subsequently, the step of providing a data set related to the type information using the second data generated in step S630 (S640) may be performed.
[0165] A specific step of attribute data according to the present invention may be performed by identifying a subspace corresponding to type information of a second content within an attribute space of predefined structural and stylistic dimensions for each content type, and determining at least one attribute value from the identified subspace. This specific step of attribute data is a computing operation different from human cognitive classification or labeling, and may be performed by a series of computing operations including index-based access to an attribute set data structure stored in memory, reference to dimension-wise distribution information of attribute values, and serialization into a form that matches the input token sequence of an artificial intelligence model.
[0166] More specifically, the attribute data can be placed together within the same input sequence as the few-shot example derived from the first data, and accordingly, the artificial intelligence model can output the second data by referencing together the context derived from the few-shot example and the structural constraints imposed by the attribute data within a single forward pass operation for the input sequence. Due to this combined structure within the input sequence, the second data having structural and stylistic characteristics that match the content type of the second content can be obtained without a separate learning operation that additionally updates the parameters of the artificial intelligence model.
[0167] In the step of receiving the first data (S610), the control unit (700) can receive the first data corresponding to the first content through various sources (e.g., a website, a database (DB), an API, a server linked to the data generation system (1000), a central server, an external server, cloud storage, a user terminal (10), a large data set, etc.).
[0168] In the present invention, “first data” refers to data corresponding to existing content, and in one embodiment, said first data may be configured in the form of a query-response pair including a query and a response to said query. For example, said first data may be data composed of at least one query and response corresponding to a specific content type.
[0169] In this case, there may be various methods for receiving the first data in the present invention. At this time, the first data may correspond to a data set containing query-response pairs corresponding to type information for the first content.
[0170] In one embodiment, as illustrated in FIG. 7, the control unit (700) may provide (or output) a service page (2000) associated with a data generation system (1000) to a user terminal (10). In this case, the control unit (700) may receive first data corresponding to the user input based on the user input received through the service page (2000).
[0171] In another embodiment, the control unit (700) may receive the first data based on the input of data corresponding to the user's selection among the data stored (or embedded) in the storage (or memory or storage space or database) of the user terminal (10).
[0172] In another embodiment, the control unit (700) may receive from the user terminal (10) link information (e.g., a URL) containing the first data or link information of an external storage service storing the first data. In this case, the control unit (700) may directly access the first data through the link information or obtain the first data by downloading it.
[0173] However, the method of receiving the first data in the present invention is not necessarily limited to the example mentioned, and it is obvious that the first data can be received through various methods other than the example mentioned.
[0174] Such first data may include query-response pairs corresponding to the first content. For example, as illustrated in FIG. 8, the first data (810) may include at least one first query corresponding to type information for the first content and first query-response pairs (811) configured to have a response to the first query. However, the query-response pairs included in the first data (810) are not necessarily limited thereto and may further include query-response pairs corresponding to at least one other content (e.g., third query-response pairs corresponding to the third content, fourth query-response pairs corresponding to the fourth content, etc.).
[0175] As another example, if the first data (810) corresponds to a data set, the data generation system (1000) can obtain refined first data by performing safety filtering and irrelevant content removal on the first data (810). Subsequently, the data generation system (1000) can sample at least some of the first query-response pairs from the refined first data and use them as few-shot examples within the prompt. In this case, the number of query-response pairs (i.e., few-shot examples) sampled can be singular (1) or plural (e.g., 1, 2, 3, 4 or more).
[0176] Additionally, the data generation unit (500) may specify (or set, determine) at least one content that is different from the received (or input) first content among a plurality of different content as the target content (or target content). This target content may also be understood as content related to the data to be generated. For example, the target content may be specified independently by the data generation system (1000) or may be specified based on a user request received through the user terminal (10). However, the present invention is not limited to either one.
[0177] The data generation unit (500) receives (or specifies) type information for the target content and can generate at least one attribute data related to the target content based on the received type information.
[0178] The above “attribute data” refers to data that reflects the structural and stylistic characteristics of the content corresponding to the type information, and as an embodiment, the attribute data may include at least some of a plurality of attributes defined according to the content type. For example, the attribute data may include characteristics such as narrative structure, rhythmic structure, character composition, and mode of expression.
[0179] For example, let us assume that the target content is a second content having type information (e.g., Movie Script) that is different from type information (e.g., Story) for the first content. As illustrated in FIG. 9, the data generation unit (500) receives type information (822, or second type information) for the second content that is different from type information (812, or first type information) for the first content, and can generate at least one attribute data (823) related to the second content based on the received type information (822).
[0180] Specifically, the data generation unit (500) can generate attribute data (823) related to the second content based on at least some of the attributes defined for the type information (822) for the second content among the attributes (or sets of attributes) defined for each of the type information for each of the multiple different contents.
[0181] The above attributes may represent structural and stylistic dimensions (or characteristics) that distinguish multiple pieces of content. These attributes may refer to core structural and stylistic dimensions that enable different content or forms of expression to be distinguished from one another. The attributes include the unique compositional methods (e.g., narrative structure, development method, formal rules, etc.) and modes of expression (e.g., tone, rhythm, style, expressive techniques, etc.) of a specific genre, and can be defined as a set of elements that serve as criteria for the creation, interpretation, and classification of each piece of content.
[0182] For example, specific content may include formal constraints such as rhythmic structure or repetitive patterns, while other content may be defined based on narrative characteristics such as the structure of character interactions or the flow of plot development. In other words, attributes can function as criteria that enable clear distinction between different types of content and consistent creation by defining the formal regularities and expressive characteristics that the content follows. This can be derived from the fact that various types of content possess different structural and functional constraints, and the uniqueness of a genre can be defined through the combination of attributes that characterize each piece of content.
[0183] As a specific example, the data generation unit (500) may select at least some of the attributes defined for type information (822) for the second content among the attributes, and generate attribute data (823) related to the second content based on the selected at least some of the attributes. In this case, the number of attributes included in the attribute data may be singular (1) or plural (e.g., 1, 2, 3, 4 or more).
[0184] That is, the data generation unit (500) may select at least some attributes from predefined attributes (823a, 823b, 823c, or set of attributes) corresponding to type information for the second content, and generate the attribute data (823) including at least some of the selected attributes. For example, the first attribute (823a) may be an attribute related to structure and pacing, the second attribute (823b) may be an attribute related to characterization and dialogue, and the third attribute (823c) may be an attribute related to visual storytelling and cinematic flow.
[0185] Furthermore, the data generation unit (500) can input at least one of the first data and attribute data into an artificial intelligence model to generate second data corresponding to type information for the second content.
[0186] The “second data” above refers to data generated to correspond to a content type different from the first data, and in one embodiment, the second data may be configured in the form of a query-response pair including a query corresponding to the second content and a response to said query. Additionally, the “type information” refers to information for distinguishing the type or genre of content, and in one embodiment, the type information may represent at least one of a plurality of distinct content types, such as a story, poetry, rap, movie script, game design, character design, podcast script, or TV series. As an example, the type information of the first content may be a movie script, and the type information of the second content may be a TV series.
[0187] To this end, the data generation unit (500) may generate second data by configuring a prompt for generating second data and inputting it into an artificial intelligence model. In this case, the artificial intelligence model may correspond to at least one of a large-scale language model (LLM) or a generative artificial intelligence (Generative AI) model.
[0188] Here, configuring a prompt can be understood as configuring input data (or input information) for generating second data, and such input data (e.g., a prompt) can be processed as input to an artificial intelligence model.
[0189] More specifically, as illustrated in FIG. 8, the data generation unit (500) can construct a prompt (815) using at least one of the first data (810, or the first query-response pairs (811)), type information (822) for the second content, and attribute data (823) related to the second content. Then, the data generation unit (500) inputs the prompt (815) into an artificial intelligence model (510) and can obtain the second data generated from the artificial intelligence model (510) based on the prompt (815). At this time, the first query-response pairs (811) may be included in the prompt (815) as few-shot examples.
[0190] As a specific example, when a prompt (815) is input into an artificial intelligence model (510), the artificial intelligence model (510) can generate a second query (821a) corresponding to type information (822) for the second content based on the prompt (815). In this case, the number of second queries (821a) generated can be singular (1) or plural (e.g., 1, 2, 3, 4 or more).
[0191] In the present invention, the prompt (815) can be configured as shown in the following example (see Table 3). In the present invention, the prompt (815) can also be described as a “structured prompt.”
[0192] Attribute-Guided Query Generation Prompt Template You are a {genre} prompt generator.{task_instruction}(Guideline): Create distinct scenarios. When crafting each prompt, naturally incorporate these specific dimensions:{genre_attribution}(0 to len(genre_attribute) distinct attributes sampled specific to the target genre. Draw thematic inspiration from the examples below, but reimagine them with new perspectives suitable for {genre}.{few_shot}(Sampled from existing Story Generation Query Pool for thematic variety.)
[0193] Next, the data generation unit (500) may proceed with the process (or step) of generating a response (821b) to a second query (821a) generated (or obtained) from the artificial intelligence model (510). In this case, the number of responses (821b) generated may vary depending on the number of second queries (821a) generated, and may be singular (1) or plural (e.g., 1, 2, 3, 4 or more).
[0194] In this case, the data generation unit (500) may use the artificial intelligence model (510) to generate a response (821b) to the second query (821a). For example, the data generation unit (500) may input the second query (821a) into the artificial intelligence model (510) to obtain a response (821b) to the second query (821a) generated (or output) from the artificial intelligence model (510).
[0195] Alternatively, the data generation unit (500) may use at least one artificial intelligence model different from the artificial intelligence model (510) to generate a response (821b) to the second query (821a). In this case, the artificial intelligence model (510) that performs the role of generating the second query (821a) in the present invention may be referred to as the “first model,” and the artificial intelligence model (not shown) that performs the role of generating a response (821b) to the second query (821a) may be referred to as the “second model.” The data generation unit (500) may input the second query (821a) generated from the first model into the second model to obtain a second response (821b) to the second query (821a) generated (or output) from the second model.
[0196] That is, in the present invention, the process of generating a query and a response to the query may be performed using a single model, or the process of generating a query and the process of generating a response to the query may be performed using multiple different models. However, the present invention is not limited to any one of these.
[0197] Subsequently, the data generation unit (500) may generate the second data using the second query (821a) and the response (821b) to the second query. For example, as illustrated in FIG. 10, the second data (820) may include second query response pairs (821) configured to have the second query (821a) and the response (821b) to the second query (821a). Generating such second data (820) can also be understood as generating data (i.e., a query-response pair) in which the second query (821a) and the response (821b) form a pair.
[0198] Meanwhile, the control unit (700) can perform a quality evaluation of the second data (820) using the evaluation unit (600).
[0199] The evaluation unit (600) can perform a quality evaluation of the second query-response pairs (821) using an evaluation model (610) that performs a quality evaluation of the generated data.
[0200] In this case, the evaluation unit (600) may configure a prompt (or evaluation prompt, quality evaluation prompt, etc.) to be used for quality evaluation of the second query-response pairs (821) in order to perform quality evaluation of the second query-response pairs (821). Such a prompt may be configured as in the following example (see Table 4).
[0201] Response Evaluation PromptSystem Prompt You are an impartial expert judge evaluating the quality of an AI assistant's response to a user prompt. You will be provided with: 1. The user's prompt 2. A reference answer (for comparison only; it is not necessarily correct or ideal) 3. The evaluated assistant's answer Goal: Assess how well the evaluated assistant's answer performs relative to the reference answer and how effectively it fulfills the user's intent. Do not assume the reference answer is correct—treat it as a baseline or comparison point to judge informativeness, reasoning quality, and clarity. Evaluation Criteria: 1.Relevance to Prompt & Theme - Does the writing align with the user's instructions, mood, tone, or narrative intent? 2. Creativity & Originality - Does the piece show imaginative expression, a unique voice, or fresh ideas beyond clichés? 3. Imagery & Language Quality - Are word choices vivid, expressive, emotionally resonant, and stylistically appropriate? 4. Coherence & Structure - Is the writing logically or aesthetically structured, with smooth transitions and clear narrative flow? 5. Language Appropriateness & Code Switching - Is the language use appropriate for the expected primary language?* If the response contains unnecessary code switching (e.g., excessive or unjustified mixing of languages), apply a strong penalty (Apply a strong penalty for unnecessary code switching (excessive or inappropriate mixing of languages)).* Technical or domain-specific terms that reasonably require another language are acceptable (Use of another language is acceptable if it is reasonable due to technical or domain-specific terminology)* Awkward calques, translation-like phrasing, or foreign sentence structures that reduce naturalness should lower the score (Awkward literal translations or foreign sentence structures that reduce naturalness should lower the score) Scoring: Provide a single overall score on a 1-10 scale reflecting the following two factors:* The absolute quality of the answer, and * Its relative improvement or inferiority compared to the reference answer) Scoring Guide:* 1-2 → Poor: unfocused, shallow, or stylistically ineffective (unfocused, superficial, and stylistically ineffective).* 3-4→Weak (3-4 → insufficient): attempts creativity but lacks clarity, expression, or thematic presence.* 5-6→Comparable: acceptable and clear, but not notably expressive or meaningful.* 7-8→Strong: more vivid, engaging, or thematically coherent than the reference.* 9-10→Exceptional: highly evocative, polished, and artistically superior to the reference. Output Requirements:* Output must be in JSON format only.* Do not include explanations or additional text.* Use the following schema:{"score": . <integer>, / overall rating from 1 to 10(overall score between 1 and 10)"code_mix": <boolean> / true if unnecessary code switching is present(true if unnecessary code switching is present)}
[0202] The evaluation unit (600) can input a prompt such as the one in Table 1 above into the evaluation model (610). Based on the prompt, the evaluation model (610) can calculate a score (or evaluation score, quality score, quality evaluation score, etc.) according to the quality evaluation of each of the second query-response pairs (821).
[0203] And, the evaluation unit (600) can select at least some of the second query-response pairs (821) that satisfy a preset standard based on a quality evaluation (830).
[0204] More specifically, the evaluation unit (600) may select at least some pairs (830) among the second query-response pairs (821) that satisfy a preset criterion based on the score for each of the calculated second query-response pairs (821). For example, the at least some pairs (830) that satisfy the preset criterion may be pairs that received a score that is 2 standard deviations or more above the mean or exceeds it. On the other hand, query-response pairs that do not satisfy the preset criterion may be pairs that received a score that is 2 standard deviations or less above the mean or less than it is. In this case, among the second query-response pairs (821), the query-response pairs that do not satisfy the preset criterion may be excluded so as not to be used in the process of generating (or constructing) the final data set. That is, they may be excluded so as not to be included in the final data set.
[0205] Furthermore, the control unit (700) may generate (or construct) a data set (900) containing at least some of the second data and provide it to the user terminal (10). That is, the control unit (700) may provide the user terminal (10) with a data set (900) containing at least some of the pairs (830) selected based on quality evaluation among the second query-response pairs (821).
[0206] The results of generating such data sets (900) can be visually provided through a user interface.
[0207] For example, as illustrated in FIG. 11, the control unit (700) may provide the result of generating a data set (900) to a service page (2000) displayed on the display unit of the user terminal (10). In this case, the result of generating the data set (900) may include the result of generating second query-response pairs (901, 902, 903, 904) corresponding to type information for the second content. Through this, the user can intuitively recognize the result of generating the data.
[0208] Meanwhile, the data set (900) generated in the present invention may include data corresponding to each of the type information for each of the different plurality of contents. These data may include query-response pairs corresponding to each of the type information.
[0209] For example, as illustrated in FIG. 12, the data set (900) generated in the present invention may include query-response pairs (911, 912, 913, 914) corresponding to each of type information for each of a plurality of contents (e.g., Podcast Script, Motivational Speech, Game Design, etc.).
[0210] In this way, the present invention provides a dataset including multiple content types by constructing data based on a set of attributes defined for each content type for various content types, and said dataset can be utilized for training an artificial intelligence model.
[0211] Meanwhile, the data generation method according to the present invention can also be represented by the algorithm shown in FIG. 14. Since more specific details regarding the data generation method have been described above, they will be briefly explained below to avoid duplication of description.
[0212] First, the data generation system (1000) can perform safety filtering and remove unnecessary content on the input data (e.g., seed data set).
[0213] Next, in the first step (e.g., Step 1), the data generation system (1000) randomly selects a plurality of query-response pairs from the refined data and uses them as few-shot examples.
[0214] Subsequently, in the second step (e.g., Step 2), the data generation system (1000) may randomly select some of the various attributes defined for the target content. At this time, the number of attributes selected may vary, and no attributes may be selected, or multiple attributes may be selected. As the number of attributes increases, the generated query may have specific constraints.
[0215] Furthermore, in the third step (e.g., Step 3), the data generation system (1000) can configure a prompt based on the few-shot example and the selected attribute, and input it into an artificial intelligence model to generate a plurality of new queries.
[0216] And, the data generation system (1000) can generate a response for each generated query and evaluate the quality of the query and the response using an evaluation model. The data generation system (1000) can include the query-response pair in the final data set only if the evaluation score satisfies a preset criterion.
[0217] Furthermore, the data generation system (1000) according to the present invention may construct a final data set (900) by additionally considering the degree of redundancy between data for the data selected through quality evaluation by the evaluation model. In one embodiment, the data generation system (1000) may calculate the degree of redundancy between the selected data and extract data in which the calculated degree of redundancy is less than a preset redundancy standard and include it in the data set (900). The degree of redundancy may be calculated based on at least one of textual similarity between data (e.g., Jaccard similarity, cosine similarity, edit distance-based similarity, n-gram redundancy), semantic similarity in an embedding space (e.g., cosine distance between embedding vectors), or structural similarity (e.g., the match rate of the query structure, the match rate of the response structure).
[0218] The above duplicate criteria may be set as an absolute similarity threshold or as a dynamic criterion based on the relative distribution of the entire selected data pool. For example, the data generation system (1000) may extract only the data among the selected data in which the mutual cosine similarity is less than a preset threshold (e.g., 0.85) and include them in the data set (900). Additionally, the data generation system (1000) may perform clustering-based duplicate removal (e.g., selecting only representative data from similar data clusters), greedy diversity extraction (e.g., Maximum Marginal Relevance), or deterministic points
[0219] Through this redundancy-based extraction process, the data set (900) can be configured to include high-quality data and diverse data with reduced mutual redundancy. That is, in the present invention, as evaluation model-based quality control and redundancy-based diversity control are combined and applied, a data set satisfying both high quality and high diversity simultaneously can be provided, which is suitable for use in artificial intelligence model training or application services.
[0220] The data generation system (1000) can repeat the process described above to finally construct a data set containing data corresponding to each of the type information for each of the multiple contents.
[0221] Meanwhile, FIGS. 15 and 16 are drawings showing performance related to a data generation method and system according to the present invention.
[0222] In one embodiment, referring to FIG. 15, it can be seen that a dataset generated using the data generation method according to the present invention achieves a balanced distribution for a plurality of target contents, and that each category maintains a sufficient proportion. These results indicate that, along with the balanced distribution, the data generation method according to the present invention successfully generates specialized queries for various contents while maintaining structural consistency within each category.
[0223] In another embodiment, referring to FIG. 16, a model trained using a dataset according to the present invention showed result values for the evaluation items shown in FIG. 16 compared to a conventional model specialized in long text generation. This may mean that a dataset containing various content can contribute to improving the model's generalization performance and creative generation ability, rather than a single content-centered learning method.
[0224] Meanwhile, the data generation method and system according to the present invention may also be applied to inputs including long text or long text context. In particular, the present invention can perform data processing that converts into different content types while maintaining long text context.
[0225] In one embodiment, the data generation system (1000) can generate data to enhance the understanding performance of the model's long text or long text context.
[0226] First, the data generation system (1000) can receive at least one first data corresponding to the first content. This first data may include a long text form composed of multiple sentences. For example, the first data may include a narrative structure of more than one paragraph, a continuous flow of explanation, or text in which multiple sentences are interconnected to form meaning.
[0227] Next, the data generation system (1000) receives type information for a target content (e.g., a second content) that is different from type information for a first content, and can generate attribute data related to said type information. At this time, the generated attribute data may include not only style information but also contextual characteristics such as the method of context development in a long text structure, information flow, and narrative structure. That is, in the present invention, different long text composition methods and context maintenance methods for each content type are defined as attributes, and based on this, an understanding of the long text context can be reflected in the subsequent data generation process.
[0228] Subsequently, the data generation system (1000) may construct a prompt using at least one of the first data and the attribute data, and input it into an artificial intelligence model to generate second data. At this time, the prompt may include a few-shot example containing long text data composed of multiple sentences, through which the model may generate second data by reflecting connectivity between sentences, flow between paragraphs, and context at the overall text level. Additionally, the generated second data may have the characteristic of maintaining semantic consistency at the overall paragraph or document level.
[0229] Furthermore, the data generation system (1000) can perform a quality evaluation of the generated second data. In this case, the quality evaluation can be performed based on contextual consistency, logical connectivity, and information retention at the paragraph or document level, as well as at the sentence level. In particular, for data containing a long text context, the retention of meaning in the overall flow and structural completeness may be included as evaluation factors, and based on the evaluation results, only data satisfying the criteria can be selected and provided as a data set.
[0230] Accordingly, the present invention can generate data to improve the understanding performance of a model regarding long text data (e.g., long texts, long text contexts), and in particular, can generate high-quality data that is contextually consistent and structurally complete for various content types. A model trained with such data has improved understanding performance regarding long texts or long text contexts, and its ability to maintain semantic consistency at the paragraph and document levels can be enhanced.
[0231] As described above, according to the data generation method and system of the present invention, by utilizing first data corresponding to the first content and attribute data related to the content to be generated, second data having a different type from the first content can be generated based on an artificial intelligence model. That is, the present invention enables the generation and expansion of data of various content types and can construct a data set encompassing various content types. Accordingly, as the difference between the type information for the first content and the type information for the second content is reflected in the attribute data generation step, the present invention allows a data set including multiple different content types to be constructed using the artificial intelligence model. Accordingly, in the present invention, a conversion from an existing content type to data corresponding to a content type different from the existing content type can be performed, and through such conversion, the data set can be utilized in various application fields.
[0232] Furthermore, according to the data generation method and system of the present invention, a prompt can be configured using at least one of the first data, type information, and attribute data, and the prompt can be input into an artificial intelligence model to generate second data. That is, the present invention can generate high-quality data that is structural and consistent for various content types, and can simultaneously ensure accuracy and diversity in data generation.
[0233] Furthermore, according to the data generation method and system of the present invention, attribute data is generated based on structural and stylistic attributes defined for each type of content information, and new data is generated using this, thereby enabling the generation of customized data that reflects the characteristics of each content type. Through this, the present invention can construct a data set including various content types.
[0234] Furthermore, the present invention can provide a data set containing high-quality data by using an evaluation model to perform a quality evaluation of the second data and selecting at least a portion of the second data based on the quality evaluation. Accordingly, the provided data set may be configured to include only pairs that satisfy the preset criteria, and pairs that do not satisfy the preset criteria may be excluded from the data set.
[0235] The quality evaluation and screening step according to the present invention is not limited to a single threshold comparison of the score calculated by the evaluation model, but may be performed in the form of a multi-stage filtering pipeline in which at least one of a range based on score distribution statistics (e.g., mean, variance, quartiles, standard deviation), weights set differently for each content type, and screening based on redundancy is applied sequentially or in parallel according to a predetermined application order. Compared to conventional methods relying on a single threshold or human inspection, this multi-stage filtering structure can improve the robustness of the evaluation signal against noise and reduce the amount of human intervention required to achieve the same quality standard.
[0236] More specifically, the multi-stage filtering pipeline may be configured to first exclude candidates that clearly fall below the standard based on a score threshold in the initial stage, and to further exclude outliers among the remaining candidates by applying statistical distribution-based screening in a subsequent stage. Due to this stepwise exclusion structure, subsequent calls to the evaluation model or the calculation of additional evaluation metrics may be performed only on the reduced set of candidates, and consequently, the cumulative number of inferences of the evaluation model required throughout the entire process of building training data may be reduced.
[0237] Furthermore, according to the data generation method and system of the present invention, by providing the generated data set through a user terminal, it supports the user in intuitively checking and utilizing data of various content types, and can improve the overall efficiency of data generation and utilization and user convenience.
[0238] The generation of the data set according to the present invention is not limited to intra-type augmentation within a single content type, but can be performed as an inter-type transformation that derives output data belonging to a second content type different from the first content type from input data belonging to a first content type. This transformation differs from a simple summary, classification, or paraphrase within the same type regarding the semantic content of the input content, and may include an operation in which the mode of expression is reconstructed according to the structural and stylistic dimensions of the content type of the output content.
[0239] The data set provided as a result of such conversion between content types may include the representational modes and structural characteristics of the second content type that are difficult to reach with a conventional artificial intelligence model trained exclusively on the first content type. Accordingly, by utilizing the data set generated by the present invention for training or fine-tuning, the model to be trained can additionally secure performance in generating or understanding the second content type while maintaining its existing performance regarding the first content type.
[0240] Prior to the proposal of the present invention, data generation methods applied in the field of content data generation primarily involved deriving new data samples within the same content type by further training an artificial intelligence model based on training data limited to a single content type, or securing data of other content types by relying on direct creation by human creators. In these conventional methods, whenever data corresponding to a new content type is secured, it may be required to collect separate training data limited to that content type, perform a separate training step, or input additional human creators. Furthermore, there may be a structural limitation in that as content types become more diversified, at least one of the training data storage volume, training computation volume, and human resource input volume may increase in proportion to the number of content types.
[0241] The data generation method and system disclosed in this specification have an information flow structure in which the first data corresponding to the first content is input into the artificial intelligence model through a first information path and utilized as a context for a representation style, and the attribute data corresponding to the content type of the second content is input into the artificial intelligence model through a second information path and utilized as a structural constraint for a generation target, and the inputs through the first information path and the second information path are integrated and processed within a single structured input information. By means of this information flow structure, the present invention can obtain data corresponding to a plurality of different content types from the same artificial intelligence model without separately updating the parameters of the artificial intelligence model for each content type.
[0242] By changing the information flow structure, the present invention reduces the need to construct separate learning models or separate learning data sets in proportion to the number of content types, and accordingly, at least one of the storage amount of learning model parameters and the cumulative storage amount of learning data may be reduced. Furthermore, the present invention can perform data generation through a pipeline centered on a single artificial intelligence model and a single evaluation model without interposing a separate conversion model or a separate processing module for conversion between content types, thereby reducing the complexity of the system configuration. Moreover, by combining sampling based on the alignment of the dimension distribution of the attribute subsets with a multi-stage filtering pipeline, the generalization performance between content types can be improved when the target model is trained on the data set, and the phenomenon where performance for other content types deteriorates due to bias toward a specific content type can be mitigated.
[0243] The information flow structure described above and the technical effects resulting therefrom can be realized by a configuration comprising, in the claims disclosed in this specification, a step of receiving the first data; a step of receiving type information for the second content; a step of specifying the attribute data corresponding to the second content based on the type information; a step of inputting input data including at least one of the first data and the attribute data into the artificial intelligence model; and a step of obtaining the second data corresponding to the content type of the second content from the artificial intelligence model; and a configuration of performing a quality evaluation of the second data using an evaluation model.
[0244] Furthermore, the present invention can achieve the technical effect of automatically generating high-quality learning data corresponding to various content types while avoiding an increase in learning resources proportional to the number of content types through specific computing operations including a sampling step based on dimension distribution alignment for a set of structural and stylistic attributes defined by content type, a structured input information configuration step that integrates a sampled attribute subset and few-shot examples within a single input sequence, and a multi-stage filtering step that combines score calculation by an evaluation model with selection based on statistical distribution.
[0245] Meanwhile, the present invention described above can be implemented based on a quantum computer. The present invention implemented based on a quantum computer may include a qubit-based quantum processor and quantum memory, and may include software and hardware interfaces optimized for quantum computation.
[0246] Quantum processors in quantum computers utilize qubits to efficiently process complex operations through parallel computation, quantum entanglement, and quantum superposition, which cannot be performed by the binary bits of classical computers. Quantum processors process data using quantum gates and can provide exponential speed improvements for specific problems.
[0247] Meanwhile, the present invention described above may be implemented as a program that is executed by one or more processors on a computer and can be stored on a medium (or recording medium) that is readable by such computer.
[0248] Furthermore, the present invention described above can be implemented as computer-readable code or instructions on a medium on which a program is recorded. That is, the present invention can be provided in the form of a program.
[0249] Meanwhile, computer-readable media include all types of recording devices in which data that can be read by a computer system is stored. Examples of computer-readable media include HDD (Hard Disk Drive), SSD (Solid State Disk), SSD (Silicon Disk Drive), ROM, RAM, CD-ROM, magnetic tape, floppy disk, optical data storage device, etc.
[0250] Furthermore, the computer-readable medium may be a server or cloud storage that includes a storage and is accessible to an electronic device via communication. In this case, the computer may download the program according to the present invention from the server or cloud storage via wired or wireless communication.
[0251] A computer program may reach the system (1000) through various suitable transmission mechanisms. The transmission mechanism may be, for example, a computer-readable storage medium, a computer program product, a memory device, a recording medium such as a CD-ROM or DVD, or a product that tangibly embodies the computer program. The transmission mechanism may be a signal configured to reliably transmit the computer program through air or an electrical connection. The system (1000) may propagate or transmit the computer program as a computer data signal.
[0252] Furthermore, references to 'computer-readable storage media,' 'computer program products,' 'computer programs embodied in a tangible form,' etc., or to 'controller,' 'computer,' 'processor,' etc., should be understood to include not only computers with various architectures such as single / multi-processor architectures and sequential (Von Neumann) / parallel architectures, but also specialized circuits such as Field-Programmable Gate Arrays (FPGAs), Application Specific Circuits (ASICs), signal processing units, and other devices. References to computer programs, instructions, code, etc., should be understood to include software for programmable processors or firmware, such as programmable content for hardware devices, whether it is instructions for a processor or configuration settings for a fixed-function device, gate array, or programmable logic device.
[0253] Furthermore, in the present invention, the computer described above is an electronic device equipped with a processor, namely a CPU (Central Processing Unit), and no special limitations are placed on its type.
[0254] Meanwhile, the above detailed description should not be interpreted restrictively in all respects but should be considered exemplary. The scope of the invention shall be determined by a reasonable interpretation of the appended claims, and all modifications within the equivalent scope of the invention are included within the scope of the invention.
[0255] The various embodiments described above may be combined to provide additional embodiments. Additional variations including such changes may be applied to embodiments of the present invention in light of the above detailed description. In general, the terms used in the following claims should not be interpreted as limiting the claims to the specific embodiments disclosed in the specification and claims, but should be interpreted as including all possible embodiments together with the full scope of equivalents to which such claims are given. Accordingly, the claims are not limited by the present disclosure.< / boolean> < / integer>
Claims
1. In a method for generating data performed by a computer, A step of receiving at least one first data corresponding to the first content; A step of receiving type information for a second content different from the first content above; Among the attributes related to at least one of the structure and style distinguishing different content, a step of specifying attribute data corresponding to the second content based on the type information; A step of inputting input data including at least one of the first data and the attribute data into an artificial intelligence model to generate data corresponding to the second content different from the first content; A step of obtaining at least one second data corresponding to a content type of the second content different from the content type of the first data from the artificial intelligence model; and A data generation method characterized by including the step of generating a data set corresponding to the second content using the second data.
2. In Paragraph 1, The step of specifying the above attribute data is, Among the above attributes, a step of sampling a subset of attributes including attributes related to the second content; and A data generation method characterized by including a step of specifying the attribute data based on the sampled attribute subset.
3. In Paragraph 2, In the step of sampling the above attribute subset, A data generation method characterized by sampling an attribute subset such that at least one few-shot example sampled from the first data and the attribute subset are included together in the input data to the artificial intelligence model.
4. In Paragraph 1, The step of inputting into the above artificial intelligence model is, A step of configuring a prompt using at least one of the first data, the type information and the attribute data for the second content; and The above includes the step of inputting the above prompt into the above artificial intelligence model, A data generation method characterized by acquiring the second data output from the artificial intelligence model based on the prompt in the step of acquiring the second data.
5. In Paragraph 4, The first data comprises first query-response pairs, each including at least one first query corresponding to type information for the first content and a response to the first query. A data generation method characterized in that the first query-response pairs are included as few-shot examples in the prompt.
6. In Paragraph 4, The step of acquiring the second data above is, A step of obtaining at least one second query corresponding to the type information for the second content from the artificial intelligence model; and A data generation method characterized by including the step of obtaining a response to the second query above.
7. In Paragraph 6, A data generation method characterized in that the second data comprises second query-response pairs including the second query and the response to the second query.
8. In Paragraph 7, The step of generating the above data is, A step of performing a quality evaluation of the second query-response pairs using an evaluation model; and A data generation method characterized by including the step of selecting at least some pairs among the second query-response pairs that satisfy a preset criterion based on the above quality evaluation.
9. In Paragraph 8, In the above screening step, Using the above evaluation model, calculate a score based on the quality evaluation of each of the above second query-response pairs, and A data generation method characterized by selecting at least some of the second query-response pairs that satisfy the preset criteria according to statistical criteria defined based on the distribution of the scores calculated above.
10. In Paragraph 8, The above statistical criteria are, A data generation method characterized by including at least one of a deviation from the mean value, percentiles, quartiles, variances, and a standard deviation-based range.
11. In Paragraph 1, The above data is, It includes data corresponding to each of the plurality of contents, including the first content and the second content, and A data generation method characterized in that the data corresponding to each of the plurality of contents includes query-response pairs corresponding to type information for each of the plurality of contents.
12. In Paragraph 1, The above-mentioned first content and other content include at least two other contents, and The steps of generating the attribute data, inputting it into the artificial intelligence model, acquiring the second data, and generating the data set are: It is performed for each of the above at least two different contents, and A data generation method characterized in that the above data set is configured to include data corresponding to each of the at least two other contents.
13. In Paragraph 12, The step of generating the above data is, Using an evaluation model, specific data satisfying a preset quality threshold is selected among the data corresponding to each of the at least two other contents, and A data generation method characterized by configuring the data set using the specific data above.
14. In Paragraph 13, The step of generating the above data is, A data generation method characterized by extracting data from among the selected specific data such that the degree of overlap between the specific data is less than a preset overlap criterion and including it in the data set.
15. In Paragraph 1, The above first data is received from a user terminal, and A data generation method characterized in that the above data set is transmitted to the user terminal as a result of content type expansion from the first content corresponding to the first data to the second content.
16. In Paragraph 15, In the above user terminal, A data generation method characterized by data corresponding to each of the first content and the second content included in the data set being visually distinguished and displayed by content type through the user interface of the user terminal.
17. In a data generation system, Memory; and It includes at least one processor connected to the memory, and The above-mentioned at least one processor is, Input data comprising at least one of at least one first data corresponding to a first content and at least one attribute data related to a second content different from the first content is input to an artificial intelligence model, and From the artificial intelligence model above, at least one second data corresponding to the second content is obtained, and A data generation system characterized by being configured to generate a data set corresponding to the second content using the second data.
18. In Paragraph 17, The above-mentioned at least one processor is, A data generation system characterized by performing a quality evaluation of the second data using an evaluation model and generating the data set based on the quality evaluation.
19. In Paragraph 17, The above attribute data is, A data generation system characterized by being generated based on an attribute representing at least one of a structure and style that distinguishes the plurality of contents, each of which is defined for each of the plurality of contents including the first content and the second content.
20. A program that is executed by one or more processors in an electronic device and can be stored on a computer-readable recording medium, The above program is, A first instruction to input input data comprising at least one of at least one first data corresponding to a first content and at least one attribute data related to a second content different from the first content into an artificial intelligence model; A second instruction to obtain at least one second data corresponding to a content type of the second content different from the content type of the first data from the artificial intelligence model; and A program stored on a computer-readable recording medium characterized by including a third instruction to generate a data set corresponding to the second content using the second data.
