Electronic device for performing function of removing artificial pattern of ai-generated text, and model training method therefor

Abstract

An electronic device for performing a function of removing an artificial pattern of AI-generated text according to the present disclosure comprises: a humanizer model for removing an artificial pattern in a document generated by an artificial intelligence (AI) model so as to output a document considered to be written by a person; a memory for storing at least one process for performing an operation of generating an interactive training dataset in order to train the humanizer model; and at least one processor for performing the operation according to the process. The at least one processor may be configured to: convert first fake text generated by AI into first real text by using a large language model (LLM) and output the first real text; construct a forward dataset including a first pair in which the first fake text matches the first real text; convert second real text written by a person into second fake text by using the large language model (LLM) and output the second fake text; construct a reverse dataset including a second pair in which the second real text matches the second fake text; and train the humanizer model by using the forward dataset and the reverse dataset.

Description

Electronic device that performs the function of removing artificial patterns from AI-generated text and its model training method

[0001] The present disclosure relates to a method for constructing text using an artificial intelligence model, and more specifically, to an electronic device that performs the function of removing artificial patterns from AI-generated text and a method for training the model thereof.

[0002] Recently, various language generation services based on advanced Large Language Models (LLMs), such as OpenAI ChatGPT and Google Gemini, have been launched. These services are generative AI services provided in diverse ways, including document creation, proofreading, translation, question answering, and information retrieval, and can be delivered by integrating generative AI models into the business logic of software services.

[0003] Recently, generative AI technology has been able to generate language appropriate to a given context through high-level language comprehension and reasoning capabilities, but it still has limitations in terms of the quality of the generated text. Although the language generated by generative AI models is understandable to humans, it is written as sentences reflecting mechanical characteristics, which has the problem of giving readers an unnatural impression.

[0004] Because the unnatural impression of a sentence arises from potentially reflected semantic elements, such as nuances, in addition to explicitly reflected expressive elements, it was very difficult to train a model or obtain training data by extracting features to generate sentences by removing mechanical features.

[0005] In conventional patent literature, candidate text features were extracted from machine-generated text and user-defined text, and training data was constructed by selecting text features from the extracted candidate text features.

[0006] However, using methods that select text features by treating machine-generated text and user-defined text separately makes it very difficult to select consistent features in texts where expressive diversity is guaranteed.

[0007] Whether AI-generated output exhibits unnatural or mechanical patterns directly impacts the quality of the results. This is particularly important when implementing AI agent services that interact with users, where it is crucial to generate text that is natural, emphasizes lexical individuality, and ensures expressive diversity.

[0008] Therefore, the quality of AI-generated text must be improved through a separate processing step, and technology is required to remove unnatural patterns and refine machine-generated text.

[0009] To solve conventional problems, the embodiments disclosed in this disclosure aim to provide an electronic device and a method for training a model thereof that perform the function of removing artificial patterns in AI-generated text, wherein text pairs constructed from a forward dataset and a reverse dataset are used as training data to learn the pattern differences between human-written and machine-written documents for a single sentence.

[0010] The problems that this disclosure aims to solve are not limited to those mentioned above, and other unmentioned problems will be clearly understood by a person skilled in the art from the description below.

[0011] An electronic device for performing the function of removing artificial patterns in AI-generated text according to the present disclosure for achieving the aforementioned technical problem comprises: a Humanizer model that removes artificial patterns within a document generated by an artificial intelligence model (AI) and outputs a document considered to have been written by a human; and a memory storing at least one process for performing an operation of generating a bidirectional learning data set to train the Humanizer model. and at least one processor that performs the operation according to the above process; wherein the at least one processor may be configured to convert and output a first fake text generated by AI into a first real text using a large language model (LLM), construct a forward dataset including a first pair in which the first fake text and the first real text are matched, convert and output a second real text written by a person into a second fake text using a large language model (LLM), construct a reverse dataset including a second pair in which the second real text and the second fake text are matched, and train the humanizer model using the forward dataset and the reverse dataset.

[0012] A method for training a model of an electronic device to remove artificial patterns of AI-generated text performed by a processor of a device according to the present disclosure for achieving the technical problem described above, wherein the method may include: a step of converting a first fake text generated by AI into a first real text and outputting it using a large language model (LLM); a step of constructing a forward dataset including a first pair in which the first fake text and the first real text are matched; a step of converting a second real text written by a person into a second fake text and outputting it using a large language model (LLM); a step of constructing a reverse dataset including a second pair in which the second real text and the second fake text are matched; and a step of training a humanizer model using the forward dataset and the reverse dataset.

[0013] In addition to this, a computer program stored on a computer-readable recording medium for implementing the present disclosure may be further provided.

[0014] In addition to this, a computer-readable recording medium for recording a computer program for implementing the present disclosure may be further provided.

[0015] According to the aforementioned means for solving the problem of the present disclosure, by constructing a forward dataset and a reverse dataset to learn patterns when converting from human-written documents to machine-written documents, patterns when converting from machine-written documents to human-written documents, and differences between human-written and machine-written documents for a single sentence, the effect of improving the performance of a humanizer model that removes artificial patterns in AI-generated text is provided.

[0016] The effects of the present disclosure are not limited to those mentioned above, and other unmentioned effects will be clearly understood by a person skilled in the art from the description below.

[0017] FIG. 1 is a simplified block diagram illustrating the configuration of an electronic device that performs the function of removing artificial patterns of AI-generated text according to the present disclosure.

[0018] FIG. 2 is an example diagram illustrating the input / output process of a humanizer model of an electronic device that performs the function of removing artificial patterns of AI-generated text according to the present disclosure.

[0019] FIG. 3 is a flowchart illustrating a model training method of an electronic device that performs the function of removing artificial patterns of AI-generated text according to the present disclosure.

[0020] FIG. 4 is a process diagram illustrating the process of constructing a training dataset of an electronic device that performs the function of removing artificial patterns of AI-generated text according to the present disclosure.

[0021] FIG. 5 is an example diagram illustrating a fake dataset and a real dataset of an electronic device that performs the function of removing artificial patterns of AI-generated text according to the present disclosure.

[0022] FIG. 6 is a process diagram illustrating the process of constructing a fixed-direction dataset of an electronic device that performs the function of removing artificial patterns of AI-generated text according to the present disclosure.

[0023] FIG. 7 is an exemplary illustration illustrating a first prompt template of an electronic device that performs the function of removing artificial patterns of AI-generated text according to the present disclosure.

[0024] FIG. 8 is a process diagram illustrating the process of constructing a reverse dataset of an electronic device that performs the function of removing artificial patterns of AI-generated text according to the present disclosure.

[0025] FIG. 9 is an exemplary illustration illustrating a second prompt template of an electronic device that performs the function of removing artificial patterns of AI-generated text according to the present disclosure.

[0026] FIG. 10 is a structural diagram illustrating an encoder and a decoder of a humanizer model of an electronic device that performs the function of removing artificial patterns of AI-generated text according to the present disclosure.

[0027] FIG. 11 is an exemplary diagram illustrating the inference process of a trained humanizer model of an electronic device that performs the function of removing artificial patterns of AI-generated text according to the present disclosure.

[0028] Throughout this disclosure, the same reference numerals denote the same components. This disclosure does not describe all elements of the embodiments, and general content in the art to which this disclosure pertains or content that overlaps between embodiments is omitted. The terms 'part, module, component, block' as used in the specification may be implemented in software or hardware, and depending on the embodiments, a plurality of 'parts, modules, components, blocks' may be implemented as a single component, or a single 'part, module, component, block' may include a plurality of components.

[0029] Throughout the specification, when a part is described as being "connected" to another part, this includes not only cases where they are directly connected but also cases where they are indirectly connected, and indirect connections include connections made via a wireless communication network.

[0030] Furthermore, when it is stated that a part "includes" a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components.

[0031] Throughout the specification, when it is stated that a component is located "on" another component, this includes not only cases where a component is in contact with another component, but also cases where another component exists between the two components.

[0032] The terms first, second, etc. are used to distinguish one component from another, and the components are not limited by the aforementioned terms.

[0033] Singular expressions include plural expressions unless there is an obvious exception in the context.

[0034] In each step, identification codes are used for convenience of explanation and do not describe the order of the steps; the steps may be performed differently from the specified order unless a specific order is clearly indicated in the context.

[0035] The operating principles and embodiments of the present disclosure will be described below with reference to the attached drawings.

[0036] In this specification, the term "device according to the present disclosure" includes all various devices capable of performing computational processing and providing results to a user. For example, the device according to the present disclosure may include all of a computer, a server device, and a portable terminal, or may be in the form of any one of these.

[0037] Here, the computer may include, for example, a notebook, desktop, laptop, tablet PC, slate PC, etc. equipped with a web browser.

[0038] The above server device is a server that processes information by communicating with an external device, and may include an application server, a computing server, a database server, a file server, a game server, a mail server, a proxy server, and a web server.

[0039] The above portable terminal may include, for example, all types of handheld-based wireless communication devices such as PCS (Personal Communication System), GSM (Global System for Mobile communications), PDC (Personal Digital Cellular), PHS (Personal Handyphone System), PDA (Personal Digital Assistant), IMT (International Mobile Telecommunication)-2000, CDMA (Code Division Multiple Access)-2000, W-CDMA (W-Code Division Multiple Access), WiBro (Wireless Broadband Internet) terminals, smartphones, etc., as well as wearable devices such as watches, rings, bracelets, anklets, necklaces, glasses, contact lenses, or head-mounted devices (HMDs).

[0040] Functions related to artificial intelligence according to the present disclosure are operated through a processor and memory. The processor may be composed of one or more processors. In this case, the one or more processors may be general-purpose processors such as CPUs, APs, and DSPs (Digital Signal Processors), graphics-dedicated processors such as GPUs and VPUs (Vision Processing Units), or artificial intelligence-dedicated processors such as NPUs. The one or more processors control the processing of input data according to predefined operation rules or artificial intelligence models stored in memory. Alternatively, if the one or more processors are artificial intelligence-dedicated processors, the artificial intelligence-dedicated processors may be designed with a hardware structure specialized for processing a specific artificial intelligence model.

[0041] The predefined rules of operation or artificial intelligence models are characterized by being created through learning. Here, being created through learning means that a predefined rules of operation or artificial intelligence models configured to perform desired characteristics (or objectives) are created by a basic artificial intelligence model being trained using multiple learning data by a learning algorithm. Such learning may be performed on the device itself where the artificial intelligence according to the present disclosure is executed, or it may be performed through a separate server and / or system. Examples of learning algorithms include supervised learning, unsupervised learning, semi-supervised learning, or reinforcement learning, but are not limited to the examples described above.

[0042] An artificial intelligence model may be composed of multiple neural network layers. Each of the multiple neural network layers has multiple weight values ​​and performs neural network operations through operations between the results of previous layers and the multiple weights. The multiple weights possessed by the multiple neural network layers can be optimized based on the learning results of the artificial intelligence model. For example, the multiple weights may be updated so that the loss value or cost value obtained by the artificial intelligence model during the learning process is reduced or minimized. Artificial neural networks may include deep neural networks (DNNs), such as Convolutional Neural Networks (CNNs), Deep Neural Networks (DNNs), Recurrent Neural Networks (RNNs), Restricted Boltzmann Machines (RBMs), Deep Belief Networks (DBNs), Bidirectional Recurrent Deep Neural Networks (BRDNNs), or Deep Q-Networks, but are not limited to the examples mentioned above.

[0043] According to an exemplary embodiment of the present disclosure, a processor can implement artificial intelligence. Artificial intelligence refers to a machine learning method based on an artificial neural network that enables a machine to learn by mimicking human biological neurons. Methodologies of artificial intelligence can be classified according to the learning method into supervised learning, where input and output data are provided together as training data and the solution (output data) to the problem (input data) is predetermined; unsupervised learning, where only input data is provided without output data and the solution (output data) to the problem (input data) is not predetermined; and reinforcement learning, where a reward is given from an external environment whenever an action is taken from the current state, and learning proceeds in a direction that maximizes such reward. In addition, artificial intelligence methodologies can be classified according to the architecture, which is the structure of the learning model. The architectures of widely used deep learning technologies can be classified into Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Transformers, and Generative Adversarial Networks (GAN).

[0044] The device and system may include an artificial intelligence model. The artificial intelligence model may be a single model or may be implemented as multiple models. The artificial intelligence model may be composed of a neural network (or artificial neural network) and may include statistical learning algorithms in machine learning and cognitive science that mimic biological neurons. A neural network may refer to a model that possesses problem-solving capabilities by having artificial neurons (nodes) that form a network through synaptic connections and change the strength of synaptic connections through learning. The neurons of a neural network may include combinations of weights or biases. A neural network may include one or more layers composed of one or more neurons or nodes. For example, the device may include an input layer, a hidden layer, and an output layer. The neural network constituting the device can infer a result (output) to be predicted from an arbitrary input by changing the weights of the neurons through learning.

[0045] The processor can create a neural network, train or learn a neural network, perform computations based on received input data, generate an information signal based on the results of the computation, or retrain the neural network. The neural network models may include, but are not limited to, various types of models such as Convolutional Neural Networks (CNN), Region with Convolutional Neural Networks (R-CNN), Region Proposal Networks (RPN), Recurrent Neural Networks (RNN), Stacking-based Deep Neural Networks (S-DNN), State-Space Dynamic Neural Networks (S-SDNN), Deconvolution Networks, Deep Belief Networks (DBN), Restructured Boltzmann Machines (RBM), Fully Convolutional Networks, Long Short-Term Memory Networks (LSTM), and Classification Networks, such as GoogleNet, AlexNet, and VGG Network. The processor may include one or more processors to perform computations according to the neural network models. For example, the neural network is a deep neural network It may include a (Deep Neural Network).

[0046] Neural networks include CNN (Convolutional Neural Network), RNN (Recurrent Neural Network), perceptron, multilayer perceptron, FF (Feed Forward), RBF (Radial Basis Function), DFF (Deep Feed Forward), LSTM (Long Short Term Memory), GRU (Gated Recurrent Unit), AE (Auto Encoder), VAE (Variational Auto) Encoder), DAE (Denoising Auto Encoder), SAE (Sparse Auto Encoder), MC (Markov Chain), HN (Hopfield Network), BM (Boltzmann Machine), RBM (Restricted Boltzmann Machine), DBN (Deep Belief Network), DCN (Deep Convolutional Network), DN (Deconvolutional Network), DCIGN (Deep Convolutional Inverse Graphics Network), GAN (Generative Adversarial Network), LSM (Liquid State Machine), ELM (Extreme Learning Machine), ESN (Echo It will be understood by a person skilled in the art that any neural network may be included, but is not limited to, State Network, Deep Residual Network, Differential Neural Computer, Neural Turning Machine, Capsule Network, Kohonen Network, and Attention Network.

[0047] According to an exemplary embodiment of the present disclosure, the processor comprises a Convolutional Neural Network (CNN) such as GoogleNet, AlexNet, VGG Network, Region with Convolutional Neural Network (R-CNN), Region Proposal Network (RPN), Recurrent Neural Network (RNN), Stacking-based Deep Neural Network (S-DNN), State-Space Dynamic Neural Network (S-SDNN), Deconvolution Network, Deep Belief Network (DBN), Restructured Boltzmann Machine (RBM), Fully Convolutional Network, Long Short-Term Memory (LSTM) Network, Classification Network, Generative Modeling, eXplainable AI, Continual AI, Representation Learning, AI for Material Design, BERT, SP-BERT, MRC / QA, Text Analysis, Dialog System, GPT-3, GPT-4 for Natural Language Processing, Visual Analytics, Visual Understanding, Video Synthesis for Vision Processing, Anomaly Detection, Prediction, Time-Series Forecasting, Optimization for ResNet Data Intelligence, Various artificial intelligence structures and algorithms, such as recommendation and data creation, may be used, but are not limited thereto. Hereinafter, embodiments of the present disclosure will be described in detail with reference to the attached drawings.

[0048] FIG. 1 is a block diagram briefly illustrating the configuration of an electronic device (1000) that performs the function of removing artificial patterns of AI-generated text according to the present disclosure.

[0049] Referring to FIG. 1, the electronic device (1000) according to the present disclosure may include an input / output module (110), a communication module (120), a memory (130), and a processor (140). Hereinafter, the electronic device (1000) is an electronic device, and the method thereof is assumed to be implemented through the electronic device (1000).

[0050] The input / output module (110) may be various interfaces or connection ports that receive user input or output information to the user. The input / output module (110) may be divided into an input module and an output module.

[0051] The input module receives user input from the user. The input module is for inputting video information (or signals), audio information (or signals), data, or information input by the user, and may include at least one of at least one camera, at least one microphone, and a user input unit. Voice data or image data collected by the input unit may be analyzed and processed into a user control command.

[0052] User input can take various forms, including key input, touch input, and voice input. Examples of input modules capable of receiving such user input include traditional keypads, keyboards, and mice; as well as touch sensors that detect user touch; microphones that receive voice signals; cameras that recognize gestures through image recognition; proximity sensors consisting of light or infrared sensors that detect user approach; motion sensors that recognize user movements using accelerometers or gyroscopes; and all other diverse forms of input means that detect or receive various types of user input. This is a comprehensive concept.

[0053] Here, the touch sensor can be implemented as a piezoelectric or capacitive touch sensor that detects touch through a touch panel or touch film attached to the display panel, or as an optical touch sensor that detects touch by an optical method. In addition, the input module may be implemented in the form of an input interface (USB port, PS / 2 port, etc.) that connects an external input device to receive user input, instead of a device that detects user input itself.

[0054] The output module can output various types of information and provide it to the user. The output module is a comprehensive concept that includes a display for outputting video, a speaker for outputting sound (and / or an amplifier connected thereto), a haptic device for generating vibration, and various other forms of output means. In addition, the output module may be implemented in the form of a port-type output interface that connects the individual output means described above.

[0055] For example, an output module in the form of a display can display text, still images, and videos. The term "display" refers to a broad concept of an image display device that includes all types of devices capable of performing image output functions, such as Liquid Crystal Displays (LCDs), Light Emitting Diode (LED) displays, Organic Light Emitting Diode (OLED) displays, Flat Panel Displays (FPDs), transparent displays, Curved Displays, flexible displays, 1D displays, holographic displays, projectors, and others. Such a display may also take the form of a touch display integrated with the touch sensor of an input module.

[0056] In other words, the input / output module (110) can receive user input or provide output to the user based on a user interface.

[0057] The communication module (120) can communicate with an external device. Accordingly, the device can transmit and receive information with an external device through the communication module. For example, the device can communicate with an external device using the communication module so that information stored and generated within the electric vehicle charging management system is shared. The communication module (120) may include, for example, at least one of a wired communication module, a wireless communication module, a short-range communication module, and a location information module.

[0058] Here, communication, that is, the transmission and reception of data, can be performed via wired or wireless means. To this end, the communication module may be composed of a wired communication module that connects to the Internet, etc., via a Local Area Network (LAN); a mobile communication module that connects to a mobile communication network via a mobile communication base station to transmit and receive data; a short-range communication module that uses a Wireless Local Area Network (WLAN) family communication method such as Wi-Fi or a Wireless Personal Area Network (WPAN) family communication method such as Bluetooth or Zigbee; a satellite communication module that uses a Global Navigation Satellite System (GNSS) such as GPS; or a combination thereof. The wireless communication technology used for communication may include Narrowband Internet of Things (NB-IoT) for low-power communication. In this case, for example, NB-IoT technology may be an example of LPWAN (Low Power Wide Area Network) technology and may be implemented according to standards such as LTE Cat (category) NB1 and / or LTE Cat NB2, but is not limited to the names mentioned above. Additionally, or generally, wireless communication technology implemented in wireless devices according to various embodiments may perform communication based on LTE-M technology. In this case, for example, LTE-M technology may be an example of LPWAN technology and may be referred to by various names such as eMTC (enhanced Machine Type Communication).For example, LTE-M technology may be implemented in at least one of various standards such as 1) LTE CAT 0, 2) LTE Cat M1, 3) LTE Cat M2, 4) LTE non-BL (non-Bandwidth Limited), 5) LTE-MTC, 6) LTE Machine Type Communication, and / or 7) LTE M, and is not limited to the names mentioned above. Additionally or generally, wireless communication technology implemented in wireless devices according to various embodiments may include at least one of ZigBee, Bluetooth, and Low Power Wide Area Network (LPWAN) for low-power communication, and is not limited to the names mentioned above. As an example, ZigBee technology can create personal area networks (PANs) related to small / low-power digital communication based on various standards such as IEEE 802.15.4 and may be referred to by various names.

[0059] The memory (130) can store various types of information. The memory can store data temporarily or semi-permanently. For example, the memory may store an operating system (OS) for operating the first device and / or the second device, data for hosting a website, or data regarding a program or application (e.g., a web application) for generating Braille. In addition, the memory may store modules in the form of computer code as described above.

[0060] Examples of memory (130) may include a hard disk drive (HDD), a solid state drive (SSD), flash memory, ROM (Read-Only Memory), and RAM (Random Access Memory). These memories may be provided as built-in or removable types.

[0061] The processor (140) controls the overall operation of the electronic device (1000). To this end, the processor (140) performs calculations and processing of various information and can control the operation of the components of the first device and / or the second device.

[0062] The processor (140) may be implemented as a computer or a similar device according to hardware, software, or a combination thereof. Hardware-wise, the processor (140) may be provided in the form of an electronic circuit that processes electrical signals to perform control functions, and software-wise, it may be provided in the form of a program that drives the hardware processor. Meanwhile, unless otherwise specifically mentioned in the following description, the operation of the first device and / or the second device may be interpreted as being performed by the control of the processor (140). That is, the modules may be interpreted as the processor (140) controlling the first device and / or the second device to perform the following operations.

[0063] The processor (140) may be implemented with a memory that stores data for an algorithm or a program that reproduces the algorithm for controlling the operation of components within the device, and at least one sub-processor (not shown) that performs the aforementioned operation using the data stored in the memory. In this case, the memory and the processor may each be implemented as separate chips. Alternatively, the memory and the processor may be implemented as a single chip.

[0064] Additionally, the processor (140) can control one or a combination of the components described above in order to implement various embodiments according to the present disclosure, which will be described in the drawings below, on the device.

[0065] Hereinafter, with reference to FIGS. 2 to 11, an electronic device (1000) that performs the function of removing artificial patterns of AI-generated text according to one embodiment of the present disclosure and a method for learning the model thereof will be described.

[0066] As shown in FIG. 2, the humanizer model (200) can receive machine-generated text as input, perform style conversion, and output human-generated text.

[0067] In this context, machine-generated text can refer to text generated by AI or text that appears machine-like, such as text converted by AI from text written by a human.

[0068] Human-written text refers to text written directly by a human, or text generated by AI to appear as if it were written by a human (Human-like text).

[0069] The humanizer model (200) can perform a kind of paraphrase to convert the input text into a natural style and output it. The input text and the output text can maintain the same meaning, and specific units, such as sentences, paragraphs, and documents, can be maintained identically.

[0070] In one embodiment, when machine-like text is input as “The restaurant offers various options including a children’s menu, so it appears to be a suitable choice for a family dining place,” the humanizer model (200) can output human-like text by changing or adding words, word order, grammar, punctuation, etc., such as “It is a family restaurant that I highly recommend because it has a separate kids’ menu and a variety of things to eat~!”

[0071] An electronic device (1000) that performs the function of removing artificial patterns in AI-generated text according to the present disclosure may include a Humanizer model (200) that removes artificial patterns within a document generated by an artificial intelligence model (AI) and outputs a document considered to be written by a human, a memory (130) that stores at least one process for performing the operation of generating a bidirectional learning data set to train the Humanizer model, and at least one processor (140) that performs the operation according to the process.

[0072] The humanizer model (200), memory (130), and at least one processor (140) may be integrated into a single device, but the humanizer model (200) may be placed on a cloud server and the memory (130) and at least one processor (140) may be placed on separate computing devices. Alternatively, the humanizer model (200), memory (130), and at least one processor (140) may all be placed on a cloud server.

[0073] As illustrated in FIG. 3, the at least one processor (140) may be configured to perform the steps of converting a first fake text generated by an artificial intelligence model (AI) into a first real text written by a human using a large language model (LLM) and outputting it (S310), constructing a forward dataset including a first pair in which the first fake text and the first real text are matched (S320), converting a second real text written by a human using a large language model (LLM) into a second fake text generated by an artificial intelligence model and outputting it (S330), constructing a reverse dataset including a second pair in which the second real text and the second fake text are matched (S340), and training the humanizer model (200) using the forward dataset and the reverse dataset (S350).

[0074] In step (S310), a first fake text that is machine-like can be input into the LLM and converted to output a first real text that looks like it was written by a human.

[0075] In step (S320), a first pair consisting of a first fake text and a first real text can be generated and stored in a database as a forward dataset.

[0076] In step (S330), a second real text that is human-like text can be input into the LLM and converted to output a second fake text that appears to be machine-written.

[0077] In step (S340), a second pair consisting of a second real text and a second fake text can be generated and stored in the database as a reverse dataset.

[0078] At this time, the forward dataset and the reverse dataset can be stored in the same database as training datasets, and the humanizer model (200) can be trained using the training dataset stored in the database.

[0079] Specifically, as illustrated in FIG. 4, in order to build a training dataset for training a humanizer model (200), at least one processor (140) can perform a forward process for generating a forward dataset and a reverse process for generating a reverse dataset.

[0080] In the forward process (S310), AI-generated data generated by AI in advance is collected, and the collected AI-generated data is formatted into a first fake text, and a first real text is generated by paraphrasing based on the formatted first fake text.

[0081] In the reverse process (S330), human-written data written by a person in advance is collected, and the collected human-written data is formatted into a second real text and converted based on the formatted second real text to generate a second fake text.

[0082] In constructing a training dataset (S320, S340), a training dataset can be constructed by storing a first pair in which a first fake text and a first real text are matched, and a second pair in which a second fake text and a second real text are matched.

[0083] In the process of training the humanizer model (200) (S350), the humanizer model (200) is trained using the training dataset, so that the trained humanizer model (200) can receive AI-generated text or text data of that style as input and output human-written text or data of that style.

[0084] In one embodiment, the humanizer model (200) may be a deep learning-based generative language model of an encoder and decoder structure or a decoder structure that paraphrases input text and outputs text of the same unit level.

[0085] The humanizer model (200) can perform supervised learning on a large dataset of text pairs in an end-to-end manner. In addition, it can model high-dimensional features through tokenization, attention mechanisms and multilayer neural networks, and apply conventional natural language processing and deep learning techniques such as weight optimization through backpropagation algorithms.

[0086] In other words, the humanizer model (200) learns to receive AI-generated text or text data of that style as input and output human-written text or data of that style, and at least one processor (140) can use both a forward data (Machine-to-Human) construction method and a reverse data (Human-to-Machine) construction method to build a training dataset for the above learning.

[0087] As illustrated in FIG. 5, AI-generated text or text data of such style generated by AI can be defined as a Fake Dataset (10). Human-written text or text data of such style written by humans can be defined as a Real Dataset (20).

[0088] Text with patterns of documents generated by AI and text with patterns of documents written by humans can be distinguished and classified into fake data and real data.

[0089] In the Machine-to-Human data construction method, the original data is actually text generated by a machine, for example, AI, and can be called the first fake data. The first fake data generated by AI can be input into a generative AI model to generate text in a style that looks like it was written by a human, and this can be called the first real data.

[0090] By matching the first fake data and the first real data<Fake, Real'> It can be called the first pair.

[0091] In addition, in the Human-to-Machine data construction method, the original data is text actually written by a human, and this can be called the second Real Data. By inputting the second Real Data into a generative AI model, text can be generated in a style that appears to have been paraphrased by a machine, for example, an AI model, and this can be called the second Fake Data.

[0092] By matching the second fake data and the second real data<Fake', Real> It can be called the second pair.

[0093] A bidirectional dataset can be constructed by integrating forward and reverse datasets, and artificial data can be obtained by processing two types of original data separately, and a training dataset can be constructed by matching pairs of original and artificial data. By diversifying the data generation mechanism, the spectrum and distribution of collected data can be expanded, and potential bias in the training data can be minimized.

[0094] As illustrated in FIG. 6, at least one processor (140) according to one embodiment of the present disclosure can perform a forward data (Machine-to-Human) construction method.

[0095] The above at least one processor (140) may be configured to generate a prompt of the first fake text by formatting the first fake text (11) based on a first prompt template that reflects user paraphrasing elements.

[0096] The prompt module (300) can create a humanized prompt template (31). Through the created prompt template (31), context information such as the persona, system instruction phrase, and prompt phrase referenced by the LLM can be defined in advance, thereby creating a prompt that specifically guides the desired characteristics even when various data is input.

[0097] The prompt module (300) receives the first fake text (11) of the original data, the fake dataset, and the humanized prompt template (31) as input, and can format the segments of the first fake text (11) according to the humanized prompt template (31). In other words, it can generate a prompt for the first fake text by embedding the first fake text (11).

[0098] Additionally, the at least one processor (140) may be configured so that the LLM (400) receives a prompt for the first fake text and outputs a text considered to be written by a person (Real') (23).

[0099] Using LLM (400), a human-written text (23) that has been paraphrased from the first fake text (11) can be obtained.

[0100] The above at least one processor (140) may be configured to detect the first Real'-Approved text (27) by filtering the input human-written text (23) according to similarity with human-written elements, with a detection model (500) that detects whether AI is generated.

[0101] The detection model (500) can determine and filter whether the paraphrased human-written text (23) has sufficient human-written elements.

[0102] For example, if the detection model (500) determines that the text considered to be written by a person (23) contains a significant number of human-written elements that can be determined to be written by a person, it is classified as the first Real-Approved text (27), and if it determines that the text considered to be written by a person (23) does not contain a significant number of human-written elements that can be determined to be written by a person, it is predicted as AI-written text (25) and can be filtered and removed.

[0103] In one embodiment, the detection model (500) may be implemented as a supervised learning model or an unsupervised learning model or an ensemble model of multiple models, and the learned detection model (500) may classify whether the generated text was generated by a machine or written by a person.

[0104] Through the process described above, the detection model (500) selects only the text that is judged to have been written by a human, matches the original first fake text (11) with the filtered paraphrased version first real text (27), and the matched first pair can be stored as a forward dataset in the database (600).

[0105] As illustrated in FIG. 7, the prompt module (300) can create a humanized prompt template (31) composed of a persona setting part, an instruction specification part, and a data instance part.

[0106] The humanization prompt template (31) can be used to control the process of converting the first fake data (11) into a human-written document (23). The humanization prompt template (31) can be set so that a human style or expression method can be sufficiently reflected during the conversion process.

[0107] As an example, it is necessary to analyze features that are prominent only in text written by humans compared to AI-generated text, for example, the sentence length may be short, the subject may be 'I', expressions revealing opinions or thoughts or reflecting emotions may be used, special characters or symbols may be used, and the endings may be ambiguous or speculative expressions such as 'feels / is considered to be', 'can be said to be', 'seems / looks like', 'looks / thinks that', or 'is highly likely to be'.

[0108] Features that are prominent only in text written by a person can be defined as human-written elements, and instructions of the humanized prompt template (31) can be constructed based on the human-written elements.

[0109] Guidelines such as text writing style, tone, and subject matter are specifically provided through the instructions reflected in the humanized prompt template (31), and human writing elements can be more clearly reflected in the output text.

[0110] Referring to Fig. 7, instructions may be given to replace words or phrases with similar words or phrases, change the order of words, or add meaningful adverbs. Additionally, instructions may be given to add emotion and personality to the text or to generate short sentences using simple language. Guidelines may be given to match the length of the given text to a similar length, use special characters or symbols as much as possible, and avoid using emojis. For output, a dependency grammar language framework may be used instead of a syntactic structure grammar.

[0111] In other words, the instructions can be composed of guidelines that reflect the human-written elements extracted above.

[0112] Meanwhile, at least one processor (140) according to one embodiment of the present disclosure can perform a reverse data (Human-to-Machine) construction method.

[0113] Second Real Data (Real) (21), which is text written by a person, can be used as the source to generate second Fake Data (Fake') (13) that is paraphrased by a machine. Text conversion can be implemented using LLM in the same way as the forward data construction method, and Real Data can be filtered using a detection model (500), but unlike the forward data construction method, the filtered Real Data can be converted into text through LLM after filtering by the detection model (500).

[0114] The above at least one processor (140) may be configured to filter the second real text (21), into which a detection model (500) detects whether AI is generated, according to the similarity with human-written elements.

[0115] As illustrated in FIG. 8, the original data, the second real text (21), is input into the detection model (500) so that real text (22) that does not contain a significant number of human-written elements is deleted, and real text (Real-Approved) (24) that contains a significant number of human-written elements can be classified as human-written.

[0116] The original data is a document written by a person, but filtering can be performed through a detection model (500) to determine whether it has sufficient human characteristics to be used for learning.

[0117] Therefore, among the second real text (21) written by a person, only the text determined to have been written by a person can be selected and classified as real text (Real-Approved) (24).

[0118] Additionally, the at least one processor (140) may be configured to receive the second real text (24) filtered by the LLM (400) and a preset second prompt template, and output the second fake text (13).

[0119] The second prompt template can be written as a paraphrase prompt template (33). The pre-configured second prompt template (33) can be configured to include paraphrase instructions that modify the sentence structure of the second real text or replace words with similar words.

[0120] At least one processor (140) can format a segment of the filtered second test (24) to a paraphrase prompt template (33) and input a prompt with the second real text (21) embedded into an LLM (400) to generate the second fake text (Fake') (13).

[0121] Instead of the original data, the second real text (21), the filtered second real text (24) and the second fake text (13) can be mapped and stored in the database (600) as a second pair.

[0122] In other words, the reverse dataset construction method is a technique that generates new text by inputting original text written by humans into an LLM, and then constructs a training dataset using the generated text as input data and the original text written by humans as output data.

[0123] Since the paraphrase prompt template (33) is applied to text written by a person, it can be configured to be less sophisticated and less detailed than the humanized prompt template (33) used in the forward dataset construction method, as it has output data with a clear correct answer.

[0124] Paraphrase prompt templates (33) can be composed mainly of instructions related to paraphrasing, such as modifying the structure of a sentence or replacing it with a similar word.

[0125] Accordingly, at least one processor (140) can be configured to learn the humanizer model (220) to converge to real text features included in the forward dataset and the reverse dataset and to move away from fake text features.

[0126] It is possible to naturally convert text styles by analyzing patterns of fake datasets and real datasets and training features of fake text and real text in a situation where various mechanisms are considered according to the direction and process of text generation.

[0127] As illustrated in FIGS. 10 and 11, the humanizer model (200) may be a deep learning-based generative language model of an encoder (210) and decoder (220) structure or a decoder (200) structure that paraphrases input text to output text of the same unit level.

[0128] In one embodiment, the encoder (210) and the decoder (220) can be trained so that when the first fake text (11) of the first pair included in the forward dataset and the second fake text (13) of the second pair included in the reverse dataset are encoded in the encoder (210) and then decoded in the decoder (220), the first real text (27) of the first pair included in the forward dataset and the second real text (21) of the second pair included in the reverse dataset can be output.

[0129] Alternatively, as an example, the decoder (220) can be trained so that when the first fake text (11) of the first pair included in the forward dataset and the second fake text (13) of the second pair included in the reverse dataset are converted, the first real text (27) of the first pair included in the forward dataset and the second real text (21) of the second pair included in the reverse dataset can be output.

[0130] The humanizer model (200) of the learned encoder (210) and decoder (220) or decoder (220) structure can output human-written text as a conversion result, which is paraphrased as “It is a family restaurant that I really want to recommend because it has a separate kids' menu and a variety of things to eat!” when a user inputs “This restaurant offers various options including a children’s menu, so it seems like a suitable choice as a place for family dining” as AI-generated text.

[0131] Through this, in service fields requiring human-like language generation, unnatural patterns in machine-generated text can be detected, removed, and refined using various mechanisms to generate and provide higher-quality sentences.

[0132] Meanwhile, the disclosed embodiments may be implemented in the form of a recording medium that stores instructions executable by a computer. The instructions may be stored in the form of program code and, when executed by a processor, may generate a program module to perform the operation of the disclosed embodiments. The recording medium may be implemented as a computer-readable recording medium.

[0133] Computer-readable recording media include all types of recording media that store instructions that can be decoded by a computer. Examples include ROM (Read Only Memory), RAM (Random Access Memory), magnetic tape, magnetic disk, flash memory, optical data storage devices, etc.

[0134] As described above, the disclosed embodiments have been explained with reference to the attached drawings. Those skilled in the art will understand that the present disclosure may be practiced in forms different from the disclosed embodiments without changing the technical spirit or essential features of the present disclosure. The disclosed embodiments are illustrative and should not be interpreted restrictively.

Claims

1. A Humanizer model that removes artificial patterns from documents generated by an Artificial Intelligence (AI) model to output a document considered to have been written by a human; A memory storing at least one process for performing an operation to generate a bidirectional training data set to train the above humanizer model; and It includes at least one processor that performs the above operation according to the above process, and The above at least one processor is, Using a Large Language Model (LLM), the first fake text generated by AI is converted into the first real text and output, and Construct a forward dataset including a first pair in which the first fake text and the first real text are matched, and Using a Large Language Model (LLM), convert human-written second real text into second fake text and output it, and Construct a reverse dataset including a second pair in which the second real text and the second fake text are matched, and Configured to train the humanizer model using the forward dataset and the reverse dataset. An electronic device that performs the function of removing artificial patterns from AI-generated text.

2. In Paragraph 1, The above humanizer model is a deep learning-based generative language model of an encoder and decoder structure or a decoder structure that paraphrases input text to output text at the same unit level. An electronic device that performs the function of removing artificial patterns from AI-generated text.

3. In Paragraph 1, The above at least one processor is, Configured to generate a prompt for the first fake text by formatting the first fake text based on a first prompt template reflecting user paraphrasing elements. An electronic device that performs the function of removing artificial patterns from AI-generated text.

4. In Paragraph 3, The above at least one processor is, The above LLM is configured to receive the prompt of the above first fake text and output text deemed to be human-written. An electronic device that performs the function of removing artificial patterns from AI-generated text.

5. In Paragraph 4, The above at least one processor is, A detection model for detecting whether AI-generated text is configured to detect the first real text by filtering the input text deemed to be human-written according to its similarity to human-written elements. An electronic device that performs the function of removing artificial patterns from AI-generated text.

6. In Paragraph 1, The above at least one processor is, A detection model that detects whether an AI-generated text is configured to filter the input second real text based on its similarity to human-written elements. An electronic device that performs the function of removing artificial patterns from AI-generated text.

7. In Paragraph 6, The above at least one processor is, The above LLM is configured to receive the filtered second real text and a preset second prompt template as input and output the second fake text. An electronic device that performs the function of removing artificial patterns from AI-generated text.

8. In Paragraph 7, The above at least one processor is, The above-mentioned preset second prompt template is configured to include paraphrase instructions that modify the sentence structure of the second real text or replace words with synonyms. An electronic device that performs the function of removing artificial patterns from AI-generated text.

9. In Paragraph 1, The above at least one processor is, The above humanizer model is configured to learn to converge to real text features included in the forward dataset and the inverse dataset and to move away from fake text features. An electronic device that performs the function of removing artificial patterns from AI-generated text.

10. A method for training a model of an electronic device to remove artificial patterns of AI-generated text performed by a processor of the device, wherein the method comprises: A step of converting the first fake text generated by AI into the first real text using a large language model (LLM) and outputting it; A step of constructing a forward dataset including a first pair in which the first fake text and the first real text are matched; A step of converting and outputting a second real text written by a human into a second fake text using a large language model (LLM); A step of constructing a reverse dataset including a second pair in which the second real text and the second fake text are matched; and A step of training the humanizer model using the forward dataset and the reverse dataset; Humanizer model training method for removing artificial patterns from AI-generated text.

Images