Author persona-based emotion-linked reading discourse system and its method

Abstract

The present invention relates to an author persona-based emotion-linked reading discourse system and method for generating an author persona to enable a user to experience the effect of reading and conversing with the author and developing thoughts through an AI persona that reflects the author's writing style and perspective while reading a book. Through this invention, an AI persona converses like a real author, allowing users to experience the content, vocabulary, and logical flow of a text from the author's perspective, thereby enhancing reading interest and immersion. Furthermore, the difficulty of questions, the level of explanation, and the tone are immediately adjusted according to the user's emotional and cognitive state, enabling personalized learning that aids understanding in confusing situations and induces expanded thinking in immersive states. Beyond simply understanding the plot, this approach facilitates users to think for themselves and form diverse perspectives through Q&A based on the author's logic or philosophy.

Description

Technology Field

[0001] The present invention relates to an author persona-based emotion-linked reading discourse system and method for generating an author persona to enable a user to experience the effect of reading and conversing with the author and developing thoughts through an AI persona that reflects the author's writing style and perspective while reading a book. Background Technology

[0002] In the field of education, emotion-based AI coaching technology is being actively researched to support learning engagement and self-directed learning by analyzing learners' emotional states in real time. There are existing examples where various sensors, such as cameras and microphones, are used to detect users' facial expressions, gaze, and voice tone, and based on this, determine their emotional state—such as engagement, confusion, or interest—to adjust the difficulty of learning content and provide feedback. However, these technologies are primarily limited to general learning situations or digital textbooks, and the development of customized AI coaching solutions specialized for reading activities is still in its early stages.

[0003] With the recent advancements in generative AI and natural language processing technologies, reading discussion platforms have emerged where AI understands reading content and applies Socratic questioning to promote learners' critical thinking. These platforms incorporate iterative dialogue techniques into the AI ​​to encourage users to ask questions about learning topics and expand their thinking, thereby aiding students in deep understanding and reflection. However, while these Socratic AI systems primarily focus on generating open-ended questions and facilitating discussions, they are relatively lacking in features that adjust the flow and difficulty of the conversation by reflecting the user's emotional state in real time.

[0004] AI systems tailored to reading are evolving to support personalized book recommendations and learning designs by analyzing readers' tastes, comprehension, and reading history. By utilizing large-scale language models to identify key themes, writing styles, and contexts, and providing recommended books that consider user preferences, they are enhancing the quality of the reading experience. However, this approach still has limitations in providing conversational feedback by linking real-time reactions and emotions during reading to integrated systems.

[0005] With the recent advancement of conversational AI technology, there is a growing number of attempts to provide personalized conversational reading experiences by having AI analyze users' real-time behavioral and emotional data. There is a demand for customized conversational systems that detect immersion and cognitive states by integrating and analyzing multi-channel data, such as facial expressions, voice, and speech rhythm, and then use this information to adjust question difficulty and explanation methods. Prior art literature

[0006] Korean Registered Patent Publication No. 10-1683960 The problem to be solved

[0007] The objective of the present invention is to provide an immersive, personalized reading experience, as if the reader were having a real-time conversation with the author, by proposing an interactive reading system that combines an author persona with user emotion linkage and a method for providing the same. means of solving the problem

[0008] To solve the above-mentioned problem, the present invention comprises a user data collection unit (100) for collecting real-time reading response data of a user, and an interactive reading server (200) for receiving data from the user data collection unit (100), processing it in real-time, and storing and analyzing the user's conversation log and reading discourse log. The interactive reading server (200) comprises an AI persona generation module (210) for generating an AI conversational subject that mimics the unique writing style, vocabulary, and logical development method of a specific author, a user state analysis module (220) for receiving real-time response data from the user data collection unit (100) and inferring the user's current emotion and cognitive state, a conversation flow control module (230) for dynamically adjusting the content and method of conversation by linking the AI ​​persona generation module (210) and the user state analysis module (220), and a reading activity management module (240) for recording all conversation content between the user and the AI ​​author persona and providing personalized feedback based thereon. It provides an interconnected reading discourse system. Effects of the invention

[0009] Through this invention, an AI persona converses like a real author, allowing users to experience the content, vocabulary, and logical flow of a text from the author's perspective, thereby enhancing reading interest and immersion. Furthermore, the difficulty of questions, the level of explanation, and the tone are immediately adjusted according to the user's emotional and cognitive state, enabling personalized learning that aids understanding in confusing situations and induces expanded thinking in immersive states. Beyond simply understanding the plot, this approach facilitates users to think for themselves and form diverse perspectives through Q&A based on the author's logic or philosophy. Brief explanation of the drawing

[0010] FIG. 1 is a schematic diagram showing the overall configuration of the present invention. FIG. 2 is a configuration diagram showing the detailed configuration of the decomposition type reading server (100) of the present invention. Specific details for implementing the invention

[0011] Throughout the entire specification below, the same reference numerals refer to the same components unless there are special circumstances. Terms with the addition of "part" used below may be implemented in software or hardware, and depending on the embodiment, it is possible for a single "part" to be implemented as a single physical or logical component, for multiple "parts" to be implemented as a single physical or logical component, or for a single "part" to be implemented as multiple physical or logical components. Throughout the specification, when it is stated that a part is connected to another part, this may mean a physical connection between the part and the other part, or an electrical connection. Furthermore, when it is stated that a part includes another part, this does not mean that another part other than the other part is excluded unless specifically stated otherwise, but means that additional parts may be included at the designer's choice. Terms such as "first," "second," etc., are used to distinguish one part from another part, and unless specifically stated otherwise, they do not imply sequential expressions. Also, singular expressions may include plural expressions unless there is an obvious exception in the context.

[0013] The present invention relates to an author persona-based emotion-linked reading discourse system and method for generating an author persona to enable a user to experience the effect of reading and conversing with the author and developing thoughts through an AI persona that reflects the author's writing style and perspective while reading a book.

[0015] Referring to FIG. 1, the present invention comprises a user data collection unit (100) and an interactive reading server (200).

[0017] The user data collection unit (100) is a hardware device that collects real-time reading response data of the user and includes a camera, microphone, or separate wearable sensor built into the user's reading terminal, such as a tablet or smartphone.

[0018] The above user data collection unit (100) collects real-time response data, such as changes in facial expressions, eye movements, voice tone and speed, and speech rhythm, through a camera, microphone, or separate wearable sensor built into a terminal (tablet, smartphone, etc.) that the user uses for reading.

[0020] The interactive reading server (200) is central computer hardware where software modules are executed and author data and conversation logs are stored.

[0021] Specifically, the interactive reading server (200) receives data from the user data collection unit (100), processes it in real time, operates the AI ​​persona generation module (210), the user state analysis module (220), and the conversation flow control module (230), and stores and analyzes the user's conversation logs and reading discourse logs.

[0023] Referring to FIG. 2, the interactive reading server (200) is configured to include an AI persona generation module (210), a user state analysis module (220), a conversation flow control module (230), and a reading activity management module (240).

[0025] The specific details of each component are as follows.

[0027] The AI ​​persona generation module (210) is a software module that generates an AI conversational subject that mimics the unique writing style, vocabulary, and logical development method of a specific author, and includes an author data analysis unit (211) and a persona modeling unit (212).

[0029] The Author Data Analysis Department (211) is intended to collect various text data such as books, interviews, and critiques to database the author's unique writing style, vocabulary, logical structure, and core ideas.

[0030] The above text data is collected from books and literary works, including published novels, poetry collections, essays, and papers, as well as content directly mentioned by the author in newspapers, magazines, online articles, and broadcast interviews, and from book reviews, literary criticism, and commentaries about the author written by authors, critics, and scholars. This data can be collected by utilizing web scraping and APIs on open access libraries (e.g., Project Gutenberg, Internet Archive), university literature archives, e-books, news and article datasets, etc.

[0031] Text data collected by the above-mentioned author data analysis unit (211) is efficiently stored in a database after undergoing a structural analysis process. To this end, images and PDFs are converted into text using OCR (Optical Character Recognition), character encoding and normalization are performed by language and format, and preprocessing is performed by removing necessary metadata, advertisements, covers, copyright notices, etc.

[0032] The preprocessed data is classified into the categories of books, interviews, and critiques, and meta-tagged with information such as speaker details, source, year, and medium. Text mining techniques (NLP) are then utilized to extract statistics on writing style, vocabulary frequency, syntactic structure, and representative sentence lengths / forms, while automatically identifying specific keywords, repetitive expressions, the use of metaphors and symbols, major topics, and ideological keywords. For example, Orwell's representative vocabulary patterns, such as "hope," "resistance," and "if," can be identified.

[0033] Next, the logical development pattern is analyzed by component (e.g., introduction-body-conclusion), and keywords associated with predefined ideologies (e.g., humanism, cynicism, etc.) are mapped. In other words, the author's main interests and philosophy are compartmentalized through keyword extraction (TF-IDF, LDA-based) and document clustering. The compartmentalized data is stored in relational and NoSQL databases by book, utterance, and characteristic, and a refined input set is constructed for the use of LLM (Large-Scale Language Modeling). That is, tags such as "Author A_Style_Characteristics" and "Author B_Ideology_Keywords" are assigned. The databased data can be retrieved in real time by linking with the Persona Modeling Unit (212), User State Analysis Module (220), etc.

[0035] The persona modeling unit (212) fine-tunes a large-scale language model (LLM) based on data databased through the writer data analysis unit (211) to build an AI writer persona model that generates a response in the style of the corresponding writer to an input user question or situation. The AI ​​persona of the AI ​​persona model can participate in the question-and-answer dynamically generated and modified by the conversation flow control module (230).

[0036] The persona modeling unit (212) first selects a pre-trained model suitable for the project purpose and resources from among large-scale language models (LLM). Then, it obtains databased text data from the author data analysis unit, and it is preferable that this data be in a state where meta-tagging and preprocessing have been completed for the author's unique writing style, vocabulary, logical development, core ideas, etc.

[0037] To construct training data and create input-output pairs that reflect the author's style (e.g., user questions and author-style answers), a tuning dataset is built around example sentences that clearly reveal the author's writing style. If necessary, various contexts and situations are added through data augmentation, and if the amount of data is insufficient, documents or sentences with similar styles are included to enhance the learning effect.

[0038] Subsequently, to perform the fine-tuning process, supervised learning-based fine-tuning is carried out by inputting the writer style dataset into an existing LLM to adjust the model's weights. Here, hyperparameters (learning rate, batch size, number of epochs, etc.) are adjusted appropriately, and it is desirable to use cross-entropy as the loss function.

[0039] Instead of training all model weights for efficient parameter fine-tuning (PEFT, LoRA, etc.), only some important parameters are tuned to reduce costs and time and maximize efficiency, and performance evaluation and overfitting prevention measures are regularly performed using a validation dataset during model training.

[0040] The fine-tuned model evaluates the reproducibility of the author's style, grammatical accuracy, contextual appropriateness, and creativity. The evaluation method ensures objectivity by combining automated indicators (e.g., BLEU, ROUGE) with expert human evaluation, and tests whether answers to user questions are naturally generated in the actual author's style.

[0041] The finely tuned AI writer persona model is deployed to a server unit to receive user input in the form of a real-time API; when a user question or situational input is received, the model generates a response using sentences and a tone that matches the writer's style.

[0042] In response generation, prompt generation parameters such as temperature and maximum tokens are adjusted to control style and variety.

[0043] For continuous updates and management, user response data and conversation logs are analyzed and utilized to improve model performance. When new author data, such as new books, is acquired, it can be added to the training data to perform periodic retraining (fine-tuning) or further fine-tuning, and individual feedback loops can be configured for user-specific customization.

[0045] In this way, the persona modeling department (212) builds an AI writer persona model that generates responses reflecting the writer's writing style and way of thinking in response to user questions by fine-tuning the selected LLM in a specific and practical manner based on the writer's unique style data databased through the writer data analysis department (211).

[0047] The user state analysis module (220) is a software module that receives real-time response data from the user data collection unit (100) and infers the user's current emotion and cognitive state, and is configured to include a real-time data processing unit (221) and an emotion and cognitive state determination unit (222).

[0049] The real-time data processing unit (221) is for receiving raw data such as user facial expression changes, eye movement, voice tone and speed, and speech rhythm collected from the user data collection unit (100).

[0051] The emotion and cognitive state determination unit (222) is intended to classify the user into one of various states, such as immersion, confusion, boredom, or interest, by utilizing a pre-learned emotion analysis model.

[0052] The state information analyzed through the above emotion and cognitive state determination unit (222) is transmitted to the conversation flow control module (230) and used for conversation adjustment.

[0054] The emotion analysis model of the emotion and cognitive state determination unit (222) is generally designed to analyze complex data such as text, facial expressions, and voice tone by combining natural language processing (NLP), computer vision, and voice signal processing technologies. Text-based emotion recognition utilizes large-scale pre-trained Transformer models such as BERT, GPT, and RoBERTa to understand context and identify emotions, while image (facial expression) analysis utilizes a CNN (convolutional neural network) based model to precisely extract facial muscle movements and classify them into emotion categories, and voice analysis utilizes a deep learning or acoustic signal processing model that extracts tone, speed, and pitch changes from voice signals.

[0055] Here, it is possible to have an advanced multi-category classification system that includes not only basic emotions (joy, sadness, anger, surprise, etc.) but also cognitive and psychological states such as immersion, confusion, boredom, and interest.

[0057] The specific process for the emotion analysis of the emotion and cognitive state determination unit (222) is as follows.

[0058] It normalizes raw input data (text, video, audio) and extracts features. For example, facial expression data is converted into fine muscle movement points, and audio is converted into vectors such as tone, pitch, and speech rate.

[0059] Text is tokenized and converted into embedding vectors to generate contextual semantic representations, image data is converted into feature maps using CNN-based neural networks, and speech data is converted into acoustic features such as spectrograms or MFCCs.

[0060] With extracted features as input, a pre-trained neural network performs classification, and the model can categorize emotions (e.g., immersion, confusion, boredom, or interest) or perform multi-label classification.

[0061] The classification results are post-processed and output along with probability values ​​corresponding to the user's state; this information is transmitted to the conversation flow control module, etc., enabling the tracking of changes in the user's emotional state and allowing for the analysis of emotional fluctuation patterns over time.

[0063] For example, in facial expression analysis, subtle movements around the eyes, the direction of the corners of the mouth, and eyebrow positions are detected using CNNs to classify 'interest' or 'boredom,' while voice tone analysis distinguishes between 'immersion' and 'confusion' states by analyzing pitch variations and speaking speed, and text input (conversation content, etc.) can be used to extract sentiment embeddings from a transformer model to classify whether the utterance is positive, negative, or belongs to a specific psychological state.

[0064] This learned emotion analysis model is applied to raw data received by the real-time data processing unit (221) to precisely classify the user's state of immersion, confusion, boredom, interest, etc. in the emotion and cognitive state determination unit, and the result is used by the conversational reading system to adjust the conversation difficulty, tone of voice, and explanation method in real time according to the user's emotion.

[0066] The conversation flow control module (230) is a software module that dynamically adjusts the content and method of conversation by linking the AI ​​persona generation module (210) and the user state analysis module (220), and includes a question-and-answer generation unit (231) and an interaction adjustment unit (232).

[0068] The question-and-answer generation unit (231) generates questions or explanations with an author style applied by reflecting the content currently being read and the characteristics of the AI ​​author persona model of the AI ​​persona generation module (210).

[0069] The content currently being read by the above question-and-answer generation unit (231) is the original text or summary, chapter and page information of the book being read by the user, and is content data provided in real-time from the reading activity management module or reading viewer being used by the user.

[0070] The above AI writer persona model is a language model that has learned the writing style, vocabulary, and way of thinking of a specific writer, generated in the persona modeling unit (212), and includes metadata regarding the writer's unique writing style patterns, symbolic expressions, and logical flow.

[0072] The process for generating questions or explanations with the author's style applied through the question-and-answer generation unit (231) is as follows.

[0074] 1. Input data collection step

[0075] To collect input data, the original text or summary of the book currently being read by the user, along with chapter and page information (provided by the reading activity management module / reading viewer), are obtained. After text tokenization, core keywords, themes, characters, and event information are extracted. For example, in the case of a scene where the protagonist begins a new journey, keywords related to the protagonist, journey, motivation, and obstacles can be extracted.

[0076] In addition, to secure contextual information, the user's reading progress, the content of the previous conversation, and information on their recent emotional / cognitive state (transmitted from the conversation flow control module) are collected. For example, if the user was in a state of 'confusion' immediately prior, the questions are adjusted to be descriptive; if they were in a state of 'immersion,' the questions are adjusted to expand their thinking.

[0078] 2. Prompt configuration steps for generating questions / descriptions

[0079] The logic for determining the question type for configuring the prompt for generating questions / descriptions can be of the following four types.

[0080] First, there is a thought-provoking question such as, “Why did the protagonist make that decision in this scene?”; second, a comprehension-checking question such as, “What does the ‘XXX’ appearing here mean?”; third, a window-expanding question such as, “If you were the protagonist, how would you act?”; and fourth, a linked learning question that encourages comparison with other works or social situations.

[0081] In addition, the logic for determining the explanation generation type includes background explanation, conceptual explanation, and contextual explanation. The background explanation connects to the author's historical and cultural background, the conceptual explanation concerns the definition of metaphors, symbols, and technical terms within the work, and the contextual explanation concerns the significance of the current scene within the overall narrative.

[0083] The question-and-answer generation unit (231) creates a prompt template that conveys the characteristics (style, vocabulary, and logical development method) of the AI ​​writer persona.

[0085] An example of a prompt template is as follows.

[0086] [Persona Settings]

[0087] You are {Author Name}. Maintaining your writing style and way of thinking, you converse with a learner who is {User State}.

[0088] [Reading Context]

[0089] Summary of what the learner is currently reading: {Key Summary of the Text}

[0090] [Generation Goal]

[0091] Create one comprehension check question and one thinking expansion question.

[0092] Use the tone and vocabulary of {Author Name} for each question, and adjust it to suit a learner in {emotional state}.

[0094] In addition, representative vocabulary / sentence patterns (e.g., metaphor frequency, sentence length, unique phrases), logical structures (e.g., inductive development, contrast, irony), and ideological keywords (e.g., humanism, anti-totalitarianism, etc.) can be configured as prompts as needed.

[0096] 3. AI Model-Based Question / Description Generation Step

[0097] As part of the AI ​​model-based question / description generation step, a Large Language Model (LLM) is invoked for prompt input; when the current reading content, persona traits, and the purpose of the question / description are entered into the prompt, a question or explanation with the author's style applied is generated through the LLM.

[0098] In this case, parameters such as temperature (creativity control), max_tokens (output length limit), and top_p (sampling range) are used to increase result reproducibility, and these values ​​can be applied differently depending on the emotional state.

[0100] 4. Result Post-processing and Quality Verification Steps

[0101] For result post-processing and quality verification, generated questions or descriptions are compared with representative sentence patterns in the author database to check stylistic similarity, and if they fail to meet the criteria, regeneration can be requested.

[0103] The questions / explanations generated through these steps are transmitted to the interaction adjustment unit (232), adjusted in real time, and reflected in the conversation flow, and the final response is stored in the conversation log record unit (241) and subsequently utilized in the report generation unit (242).

[0105] The interaction adjustment unit (232) performs the function of generating control signals that adjust in real time the difficulty of the questions generated by the question-answer generation unit (231), the level of detail of the explanation, and the tone of the response, based on the results of the analysis of emotions and cognitive states received from the user state analysis module. The control signals are transmitted to the question-answer generation unit (231), and the generated questions and explanations are adjusted to be suitable for the user's current state.

[0107] The above interaction adjustment unit (232) may include a state mapping unit, a control rule calculation unit, and a real-time parameter generation unit.

[0108] The state mapping section stores a rule table or state-behavior mapping model that maps question difficulty, explanation detail level, and response tone to emotion / cognitive states.

[0109] The control rule operation unit derives optimal difficulty, detail, and tone values ​​by querying the rule table or executing a machine learning-based decision model based on input user state values ​​(e.g., immersion, confusion, boredom, interest, etc.). The control rule operation unit can calculate adjustment values ​​for the user's state by selectively using a rule table-based method or a machine learning and deep learning model-based method, or by using both in combination.

[0110] The real-time parameter generation unit converts the derived adjustment value into control parameters (e.g., vocabulary difficulty, sentence length, inclusion of background explanation, tone tag, etc.) of the LLM prompt or question-answer generation algorithm and outputs them.

[0112] The interaction adjustment unit (232) can operate according to the following steps.

[0113] 1) Receive user state data (emotional state and cognitive state) from a user state analysis module. The data may include state values ​​such as 'immersion', 'confusion', 'boredom', 'interest', and probability values ​​or confidence scores of the corresponding states.

[0114] 2) Refer to the state-adjustment rule table stored in the state mapping section to determine the question difficulty, level of explanation detail, and tone type corresponding to the state.

[0115] For example, the "Flow State" can be determined as "Difficulty: High, Detail: Concise, Tone: Inquiry-inducing," the "Confused State" as "Difficulty: Low, Detail: Detailed, Tone: Kind and encouraging," the "Boredom State" as "Difficulty: Medium, Detail: Includes metaphors and examples, Tone: Energetic and interest-inducing," and the "Interest State" as "Difficulty: Maintain or increase, Detail: Core-focused, Tone: Positive and conversational."

[0116] 3) The control rule operation unit fine-tunes the adjustment value by additionally reflecting user's recent conversation context information (previous conversation topic, progress, response speed, etc.). Through this, even with the same emotional state, the difficulty or level of detail can be dynamically varied according to the context of the conversation flow.

[0117] Specifically, user's recent conversation context information can be reflected by multiplying each parameter by a weight, and even with the same state value, the difficulty or detail level can be raised or lowered depending on context conditions (conversation length, response delay time, etc.).

[0118] 4) The real-time parameter generation unit converts the determined difficulty, detail, and tone values ​​into control parameters that the LLM or question-and-answer generation unit can understand.

[0119] In this case, the difficulty value is determined by vocabulary level, logical depth, and question type; the detail value is determined by whether background explanations are included and the number of additional examples; and the tone value is determined by the tone, such as 'encouraging,' 'inquiry-inducing,' or 'neutral.'

[0120] For example, in the case of ‘Difficulty: High, Detail: Concise, Tone: Inquiry-inducing’, control strings such as “Use more advanced vocabulary and analytical logical reasoning, keep sentences concise, and answer in a tone that encourages the other person to think for themselves” can be included in the LLM prompt.

[0122] 5) The above control parameters are transmitted in real time to the question and answer generation unit (231) and are directly reflected in the style and composition of the questions and explanations output by the generation unit.

[0124] In this way, the question-answer generation unit (231) reflects the control parameters transmitted from the interaction adjustment unit (232) in the prompt preprocessing stage, thereby immediately adjusting the vocabulary selection, sentence structure design, explanation length, etc., within the question / explanation generation logic.

[0126] Through this, questions or explanations with difficulty levels and tones suitable for the user's current psychological and cognitive state can be provided in real time, enabling a customized interactive reading experience tailored to specific purposes, such as maintaining immersion, resolving confusion, and stimulating interest.

[0128] The reading activity management module (240) is a software module that records all conversation content between the user and the AI ​​author persona and provides personalized feedback based on this, and includes a conversation log recording unit (241) and a report generation unit (242).

[0130] The conversation log recorder (241) is intended to store the user's questions, the AI's responses, and the user's emotions and cognitive state at that time in chronological order as a 'reading conversation log'.

[0131] The above conversation log record unit (241) performs the function of storing conversation data and user state data generated during the interaction process between the user and the AI ​​writer persona in chronological order as a 'reading conversation log'.

[0132] The above log is subsequently used in the report generation unit (242) for reading feedback, comprehension analysis, and support for expanding thinking.

[0134] In addition, the conversation log recording unit (241) may include a data receiving unit, a data linkage unit, a log storage unit, and a quality inspection unit.

[0135] The Data Receiving Unit is a module that receives user questions, AI responses, and user status analysis results in real time. The Data Linking Unit is a module that integrates questions, responses, and user status at the same conversation point into a single record. The Log Storage Unit is a module that stores generated records by session and in chronological order. The Quality Inspection Unit is a module that verifies data omissions and format consistency.

[0137] The conversation log recorder (241) can operate according to the following procedure.

[0138] 1) Data reception step

[0139] The user question text and input time are received from the question-answer generation unit (231) or the STT (speech recognition) module. Additionally, the response text generated from the AI ​​writer persona model of the persona modeling unit (212) and the response generation time are received.

[0140] Next, information on the emotional state and cognitive state at that time (immersion, confusion, boredom, interest, etc.) and the corresponding probability value or confidence score are received from the user state analysis module (220).

[0142] 2) Data Linkage and Structuring Phase

[0143] Question, response, and user status data from the same conversation point are grouped and structured into a single record. Each record includes metadata such as timestamps, session IDs, user IDs, book IDs, and chapter / page information.

[0145] 3) Log saving step

[0146] Structured records are stored in the database (DB) by session and in chronological order.

[0147] Data can be stored using relational databases or NoSQL, and saved in JSON format to facilitate subsequent analysis. Additionally, data loss can be prevented by storing data in duplicate in both local memory and the server database.

[0149] 4) Quality inspection stage

[0150] Verify whether there is missing data or inconsistency in the stored log. If missing data occurs due to network errors or recognition failures, request retransmission or mark NULL and perform subsequent correction.

[0152] S5) Subsequent module linkage step

[0153] The completed reading discourse log is transmitted to the report generation unit (242) and is used to analyze user response patterns by specific topic / chapter, track changes in understanding and emotion, and generate customized feedback and reports.

[0155] This conversation log record (241) does not simply store conversations but continuously records speech, emotions, and cognitive states along a time axis and stores them in a structured and analyzable form along with metadata, so that the entire process of reading activities can be precisely reproduced and directly utilized to produce customized feedback for learners.

[0157] The report generation unit (242) analyzes the stored logs and automatically generates a reading report that summarizes changes in understanding, changes in perspective, and key interactions regarding a specific topic, and provides it to the user.

[0159] The above report generation unit (242) performs the function of analyzing the 'reading discourse log' stored in the conversation log record unit (241), automatically generating a 'reading report' that summarizes the user's change in understanding, change in perspective, and key interaction content regarding a specific topic, and providing it to the user.

[0161] Additionally, the report generation unit (242) may include a log analysis unit, an understanding / perspective change evaluation unit, a summary and writing unit, and a report output unit.

[0162] The log analysis unit is a module that retrieves reading discourse log data by session and topic from the conversation log record unit (241) and performs preprocessing and analysis. The comprehension / perspective change evaluation unit is a module that calculates comprehension and perspective change indicators over time from the analyzed logs. The summary and writing unit is a module that automatically writes the text of a reading report based on the analysis results. The report output unit is a module that provides the written report in various formats, such as screen display, PDF / document file, or email sending.

[0164] The operation process of the report generation unit (242) is as follows.

[0165] 1) Log data collection and preprocessing step

[0166] Log data for each reading session is collected from the conversation log recorder (241). The log includes user questions, AI responses, emotional and cognitive states at that time, and metadata (book title, chapter / page, time information).

[0167] The collected data undergoes preprocessing, such as text normalization, removal of unnecessary data, and keyword extraction.

[0169] 2) Topic Classification and Key Interaction Extraction Step

[0170] Logs are classified by topic and chapter using Natural Language Processing (NLP) techniques.

[0171] Candidate summary points are generated by identifying the user's key questions, the AI's core answers, and response patterns for each segment.

[0173] 3) Analysis of changes in understanding stage

[0174] Compares changes in question difficulty, response comprehension, and emotional state between the beginning and end of the learning process.

[0175] For example, this could include the transition rate from 'confusion' to 'immersion,' or improvements in the matching rate of questions and answers regarding specific concepts. Additionally, it calculates a quantified percentage of change in understanding.

[0177] 4) Perspective change analysis stage

[0178] Analyzes changes in positive / negative / neutral attitudes and the emergence of new perspectives in utterances regarding the same topic.

[0179] At this time, the degree of semantic change relative to the initial utterance is calculated using TF-IDF, sentence embedding comparison, etc.

[0181] 5) Automatic Report Document Generation Step

[0182] Write a reading report based on the analyzed topic summary, changes in understanding, changes in perspective, and key interactions.

[0183] The report may include a summary of key content by book and chapter, a summary of major questions and answers, a record of changes in emotional and cognitive states, an analysis of changes in understanding, an analysis of changes in perspective, and recommendations for future learning and thinking expansion.

[0185] 6) Report provision step

[0186] The completed report is linked to the user profile and provided via mobile app, web, email, PDF, etc.

[0188] This report generation unit (242) quantifies and displays the user's change in understanding of the content read, records the process of expanding and changing perspectives, and provides key conversation and reaction patterns in a document, thereby enabling the user to reflect on themselves and verify learning outcomes.

[0190] The reading perspective shift tracking module (243) according to the present invention operates as a sub-component of the reading activity management module (240) by being positioned between the conversation log recording unit (241) and the report generation unit (242). This module maps all interaction records, such as utterances, questions, and responses during the user's reading, to a knowledge graph composed of semantic elements such as core concepts, ideas, people, and events extracted from the reading content, and precisely tracks the perspective shift path according to the user's conversation flow in chronological order. Through this tracking, the module quantitatively derives and accurately analyzes the user's degree of thought expansion and perspective shift patterns, and provides the results of this analysis to the report generation unit to be reflected in learning feedback, self-reflection, and the preparation of reading strategies.

[0192] The core role of the above-mentioned reading perspective movement tracking module (243) begins with identifying which node (meaning unit) within the knowledge network corresponds to an utterance in the conversation log. To this end, the module utilizes the following key factors.

[0194] First, the node set (V) is defined as a set of key concepts, figures, events, ideas, etc., automatically extracted from the reading text. For example, 'democracy', 'totalitarianism', 'truth', and 'socialism' are representative examples. These are used to minimize the unit of meaning within the text and to clarify which conceptual domain the user's utterance took place in. The above node set is formed by natural language processing (NLP) algorithms, such as Named Entity Recognition (NER), keyword extraction, and relationship extraction, analyzing the text data of the reading content to identify key elements and define them as nodes, and the values ​​of the node set are generated directly from the text data of the reading target.

[0196] Second, the edge set (E) represents semantic, logical, and contextual relationships between nodes and includes various types of relationships such as frequency of occurrence, narrative connectivity, and logical contrast. These edges are essential graph structural elements for calculating the distance of perspective shift. The edge set is extracted by various NLP techniques, such as the coexistence frequency of words and concepts within the text, connectivity within sentences or paragraphs, logical contrast, and causal relationships. It can be obtained by determining the relationship between two nodes in reading material through an NLP relationship extraction engine.

[0198] Third, the perspective shift pair (v_i, v_j) indicates which node each user's utterances are mapped to in succession over time, and this serves as a basis for tracking how semantic units are switched in the actual conversation flow. The perspective shift pair is generated in the conversation log recorder (241) based on conversation logs and user input data, within the conversation log recorder or the reading perspective shift tracking module, and is produced during the process in which the conversation log recorder (241) analyzes real-time or stored conversation utterances in chronological order to determine which node each utterance is mapped to.

[0200] Fourth, the perspective shift distance value is defined as the number of steps (unit: hop) of the shortest path on the graph between adjacent pairs of nodes at each viewpoint; a larger value indicates greater expandability of the user's thinking, while a smaller value signifies in-depth exploration within a single topic. The perspective shift distance value is calculated for the consecutive pairs of nodes generated above using a shortest path search algorithm (e.g., Dijkstra's algorithm, etc.) in a knowledge network graph pre-constructed by the reading perspective shift tracking module. In other words, it is obtained by calculating the connection distance (number of steps) between nodes within the graph structure, and this process is automatically performed by the graph search function within the module.

[0202] The reading perspective movement tracking module (243) calculates analysis indicators such as the average perspective movement distance, the repetition rate of the same node, and the frequency of passing through the central node. Here, the central node is a core concept with high connectivity within the graph, and if the user passes through it frequently, it indicates a high level of understanding of the topic structure. These multifaceted indicators are used as key judgment criteria for analyzing the user's conversation patterns and classifying modes (thought expansion mode, intensive inquiry mode, balanced inquiry mode). Furthermore, the results are ultimately reproduced in the form of visual charts and statistics in the report generation section to contribute to providing learning directions and promoting self-reflection.

[0204] The specific operation procedure of the reading perspective movement tracking module (243) is as follows.

[0206] 1) In the text analysis stage, concepts, people, events, ideas, etc. are identified in the text to be read using name recognition (NER) and relationship extraction algorithms, and then a comprehensive knowledge network is constructed by setting them as nodes (V) and edges (E), respectively. This stage is an essential process for the systematic structuring of text semantic units.

[0208] 2) In the user utterance mapping step, key keywords included in each question and response are extracted from the conversation log, and nodes within the knowledge network to which those keywords belong are searched to connect the utterances to their corresponding nodes. This process is an important mapping procedure that links conversation history with semantic units one-to-one.

[0210] 3) In the movement path calculation step, adjacent node pairs (v_i, v_j) between consecutive utterances over time are automatically generated, and the shortest path length within the graph for each pair is calculated to derive a perspective movement distance value. This value objectively defines the semantic distance between utterances.

[0212] 4) In the perspective movement volume and pattern analysis stage, the total movement volume is calculated by accumulating the perspective movement distance values ​​measured during the entire reading session, and the conversation pattern characteristics are evaluated from various angles by analyzing the average movement distance, the same node repetition rate, and the number of times the core node (central node) is passed together.

[0213] Here, the central nodes are designated from the top K% of nodes with high connectivity during knowledge network analysis. To select them, the degree and betweenness centrality of each node in the knowledge network are calculated, and the top 10% of nodes can be designated as central nodes by converting these into a comprehensive score.

[0215] 5) In the condition judgment and classification stage, if the total movement volume is above a pre-set threshold and the central node passage rate is above the threshold, it is classified as 'Accident Expansion Mode'; if the movement volume is below the threshold but the same node repetition rate is high, it is classified as 'Intensive Exploration Mode'; and otherwise, it is classified as 'Balanced Exploration Mode'. When combining conditions, the total movement volume and the central node pass rate are given equal weight, while the same node repetition rate is given a 1.2x weight to calculate the composite indicator. These threshold values ​​are initially set based on experimental and operational data and are automatically readjusted during actual operation according to user responses and learning performance.

[0216] Here, weight values ​​are typically initially set based on relevant user data, average values ​​derived from experiments, and statistical analysis. For example, equal weights are applied to the total movement volume and the central node pass rate after it is confirmed that the variable importance of the two is statistically similar by collecting and analyzing perspective shift analysis data from various actual reading sessions. Additionally, a weight of 1.2 is determined for the same node repetition rate, reflecting the fact that it yields a meaningful value despite having a relatively lower contribution rate within the evaluation metric.

[0218] 6) In the result saving and report reflection stage, numerical and pattern data such as the total sum of perspective shifts per session, the interest event group expansion index, and the in-depth inquiry index calculated by the module are stored in the database and transmitted to the report generation unit to be used as visualization and learning feedback elements.

[0220] Here, initial threshold values ​​(e.g., total movement 5 hops, central node pass rate 40%, same-node repetition rate 70%) are set as initial configuration values ​​based on the results of pilot tests and analysis of the existing user group. For weighting, equal weighting is applied to total movement and central node pass rates as their variable importance is similar in statistical analysis, while the same-node repetition rate is weighted by 1.2x to give it significance as an indicator of learning intensity. The update cycle is set to automatically adjust monthly based on a comprehensive analysis of the last 10 sessions, changes in comprehension scores, and user feedback; furthermore, the threshold values ​​are configured to be lowered if indicators consistently fall below the standard three or more times, and raised if they exceed the average by more than 20%.

[0222] In addition, the mode determination result produced by the reading perspective movement tracking module (243) is immediately reflected in reading flow control and learning recommendations. The report generation unit can present questions that stimulate in-depth analysis / critical thinking as a follow-up step, recommend comparison / contrast type tasks that connect new topics with existing topics, present questions that expand related topics, and recommend reading lists and materials to enable expansion into related concepts. Furthermore, it can present a complex learning path to ensure a balance between topic expansion and in-depth exploration.

[0224] For example, in a reading session, when a user first uttered the concept of 'democracy,' this utterance was mapped to Node 1 within the knowledge network, the second utterance, 'totalitarianism,' to Node 5, and the third utterance, 'truth,' to Node 3. The shortest path within the graph between Node 1 and Node 5 was three steps, while Node 5 and Node 3 were two steps apart. By summing the steps for each segment, the total movement volume was recorded as five steps. Furthermore, the proportion of passing through central nodes—key concepts along the path—reached half, confirming that thinking expanded within the scope of core ideas. When compared to the set threshold, the session was classified as 'Thought Expansion Mode' because the total movement volume was five steps or more and the central node passing rate exceeded 40%. This result is interpreted as a case where the user performed conscious exploration between concepts with low correlation and expanded their thinking by traversing multiple topics.

[0226] Unlike existing technologies, the reading perspective shift tracking module (243) is characterized by converting reading conversation logs into a knowledge network graph structure and quantifying perspective shift paths using multidimensional indicators such as distance, centrality, and frequency, thereby quantitatively evaluating the movement patterns within the user's thought network. As a result, the report may include an index of thought expandability beyond simple content comprehension and patterns of multi-faceted perspective shifting.

[0228] The interactive reading server (200) of the present invention further includes a user immersion and learning progress optimization module (250), and the user immersion and learning progress optimization module (250) is a component that collects learning behavior data such as emotional indicators and cognitive indicators transmitted from the user state analysis module (220), as well as past conversation logs, response speed, and answer accuracy, in real time, integrates and analyzes them, and calculates and provides conversation control variables so that the user can continuously maintain a high state of immersion and an appropriate learning progress rate during reading activities.

[0229] This module aims to minimize unfavorable learning situations, such as reduced immersion, information overload, and excessive increase in difficulty, by determining the optimal combination of state analysis results and control variables during the reading flow.

[0231] Here, the user state analysis module (220) collects state data in real time including at least one of the user's fixed gaze rate, screen switching frequency, facial expression change stability, speech duration, speech intonation change, speech length, number of follow-up questions, and answer delay time.

[0233] The above user immersion and learning progress optimization module (250) uses immersion (E), interest (H), and confusion (C) as key input factors.

[0235] The immersion index is a value normalized between 0 and 1 by combining elements such as fixation rate, screen transition frequency, facial expression stability, and speech duration, and is measured through camera-based eye tracking, facial expression recognition, and voice analysis.

[0236] The interest index is calculated as a value between 0 and 1 by analyzing the rate of change in speech intonation, the length of specific topic speech, and the frequency of positive facial expressions, and is extracted through acoustic signal analysis and facial expression and gesture recognition.

[0237] The confusion index is calculated as a value in the range of 0 and 1 by combining the frequency of re-questions, the ratio of incomprehensible responses, and the response delay time, and is obtained through question-and-answer logs and response time measurements.

[0239] To explain the additional factors, the question difficulty (Q) is set in the range of 1 to 10, where 1 is the easiest level and 10 is the most difficult level.

[0240] The depth of explanation (D) is divided into three levels: Level 1, which focuses on key summaries; Level 2, which includes background and some examples; and Level 3, which includes history, theory, and additional analogies.

[0241] Conversation tempo (S) refers to the speed of speech or the time interval between questions and answers, and can be adjusted to be fast or slow depending on the state of immersion.

[0243] The direction of control changes as the value of each factor changes. For example, if immersion and interest are high and confusion is low, the difficulty of questions is increased, explanations are made concise, and the conversation tempo is accelerated to induce the expansion of thought. Conversely, if immersion is low or confusion is high, the difficulty of questions is lowered, explanations are deepened, and the conversation speed is relaxed to restore understanding. If interest temporarily drops, analogies related to the topic are added and the tempo is adjusted to encourage participation, and if confusion rises above a certain threshold, re-explanations centered on examples are expanded to aid user understanding.

[0245] The user immersion and learning progress optimization module (250) operates through the following specific procedures.

[0246] First, during the data collection phase, data is collected in real-time via cameras, microphones, and sensors to measure immersion, interest, and confusion indicators, while response speed, whether a follow-up question was asked, and speech length are merged from previous conversation logs.

[0247] Second, in the normalization and preprocessing step, the collected raw values ​​are normalized to a range of 0 to 1, and if there are missing values, they are supplemented through interpolation.

[0248] Third, in the state judgment step, the state is classified according to the combination of immersion and confusion. For example, if the immersion level E is 0.7 or higher and the confusion level C is 0.3 or lower, it is judged as 'Challenge State'; if the immersion level E is less than 0.6 or the confusion level C is 0.5 or higher, it is judged as 'Support Needed State'; otherwise, it is judged as 'Maintenance State'.

[0249] Fourth, in the control variable calculation stage, the question difficulty, explanation depth, and conversation tempo are adjusted according to the state determination result. In the 'Challenge State', the difficulty is increased, the explanation depth is reduced, and the conversation tempo is accelerated; in the 'Support Needed State', the difficulty is lowered, the explanation depth is deepened, and the conversation tempo is relaxed; and in the 'Maintain State', the existing values ​​are maintained or slightly adjusted.

[0250] Fifth, in the condition combination and weight application step, equal weights are given to immersion and interest, and a weight of 1.5 is given to confusion to calculate a comprehensive score, and if the value is 0.70 or higher, it is set to 'Advanced Progress Mode', and if it is lower, it is set to 'Support Mode'.

[0251] Finally, in the output and reflection stage, the calculated control variable is transmitted to the conversation flow control module (230) and immediately reflected when the next turn's conversation is generated.

[0252] That is, the calculated dialogue control variables are transmitted to the dialogue flow control module (230) to adjust the difficulty level, level of detail, and speaking speed of the questions and explanations output during reading.

[0254] As a practical application example of this module, in the first case, when immersion is measured at 0.72, interest at 0.65, and confusion at 0.25, the question difficulty is raised to 7, the explanation depth is set to 2, and the conversation tempo is set to normal speed, providing challenging questions and intermediate explanations. In the second case, when immersion is measured at 0.55, interest at 0.40, and confusion at 0.55, the difficulty is lowered to 4, the explanation depth is set to 3, and the conversation tempo is adjusted to slow, resulting in in-depth explanations and relaxed conversation. In the third case, when immersion is 0.88, interest at 0.80, and confusion at 0.15, the difficulty is raised to 9, the explanation depth is set to 1, and the conversation tempo is set fast, presenting concise, high-difficulty questions.

[0256] The reference values ​​used in this module (e.g., E=0.7, C=0.3, etc.) are set based on average values ​​from pilot tests during the initial stages of system operation, and can be automatically adjusted during operation based on user feedback and performance data (understanding improvement rate, session retention rate, etc.).

[0257] This method enables more accurate and precise conversational control than existing single-indicator-based difficulty adjustment methods, preventing excessive difficulty increases in chaotic situations and facilitating the expansion of thinking through immediate in-depth questions in immersive situations.

[0259] The author persona-based emotion-linked reading discourse method consists of a data collection stage (S1), a state analysis stage (S2), a conversation flow control stage (S3), and a reading activity management stage (S4).

[0261] The data collection step (S1) is a step of collecting emotions and cognitive responses in real time through a user data collection unit (100) and transmitting them to a user state analysis module (220).

[0263] The above data collection step (S1) is a step of sampling raw data such as facial expression changes (forehead / corner of mouth movement), eye tracking coordinates, voice waveform / pitch / velocity, heart rate / skin conductivity, etc., from the camera, microphone, and wearable sensor of the user data collection unit (100) at intervals of 0.1 to 1 second and transmitting them to the user state analysis module (220) in a real-time streaming manner. When transmitting, the data is packaged in a lightweight JSON or Protobuf format, and TLS encryption may be applied to enhance security.

[0265] The state analysis step (S2) is a step in which data transmitted to the user state analysis module (220) is analyzed to determine the user's current psychological and cognitive state, and the analyzed content is transmitted to the conversation flow control module (230).

[0267] The above state analysis step (S2) performs preprocessing on video, voice, and biometric data transmitted by the real-time data processing unit (221), and then the emotion / cognitive state determination unit (222) applies a multimodal emotion analysis model (Facial Emotion CNN + voice emotion discrimination deep learning + LSTM-based sequence analysis) to classify the user's current state into one of 'immersion', 'confusion', 'boredom', or 'interest'. The classification result includes a state name, confidence level (0~1), and analysis time, and is transmitted to the conversation flow control module (230) in JSON format.

[0269] The conversation flow control step (S3) is a step of implementing a conversation subject that reflects the author's unique writing style and way of thinking through the AI ​​persona generation module (210), and generating and adjusting appropriate conversation content in real time by reflecting user state information through the conversation flow control module (230).

[0271] The above conversation flow control step (S3) configures a prompt so that the finely tuned writer persona model from the AI ​​persona generation module (210) reflects the input conversation context and the user's emotional / cognitive state, and then generates a question or explanation reflecting the writer's style through the question-answer generation unit (231). The interaction adjustment unit (232) raises or lowers the difficulty level and expands or reduces the level of detail of the explanation according to the result of the state analysis step (S2), and designates the tone of voice as one of 'encouraging type', 'inquiry-inducing type', or 'neutral type'. The final output is generated in text form, and then the TTS module converts it into speech and provides it to the user.

[0273] The reading activity management stage (S4) is a stage in which the interaction log generated by the conversation flow control module (230) is recorded and analyzed through the reading activity management module (240) to provide a reading report. By providing a reading report in this way, user immersion and learning effects are enhanced.

[0275] The above reading activity management step (S4) analyzes the 'question-answer-user status' data recorded in chronological order by the conversation log recorder (241) for each session, and then the report generation unit (242) automatically generates a reading report including a key summary by chapter, a graph of changes in user emotion / understanding, highlights of key interactions, and recommendations for future learning. The generated report is provided in the form of a PDF or an interactive screen and can be reflected in the adjustment of question difficulty in the conversation flow control step (S3) later.

[0277] Although various embodiments of the author persona-based emotion-linked reading discourse system and method have been described above, the author persona-based emotion-linked reading discourse system and method are not limited solely to the embodiments described above. Various devices or methods that can be implemented by a person of ordinary skill in the art by modifying and varying the embodiments described above may also serve as examples of the author persona-based emotion-linked reading discourse system and method described above. For example, even if the described techniques are performed in a different order than the described method, and / or the components of the described system, structure, device, circuit, etc. are combined or assembled in a form different from the described method, or are replaced or substituted by other components or equivalents, they may still serve as an embodiment of the advertising execution information providing device, system, and method described above. Explanation of the symbols

[0278] User data collection unit (100) Interactive reading server (200) AI Persona Generation Module (210) Author Data Analysis Department (211) Persona Modeling Department (212) User state analysis module (220) Real-time data processing unit (221) Emotion and cognitive state determination unit (222) Conversation flow control module (230) Question and Answer Generation Unit (231) Interaction Coordination Unit (232) Reading activity management module (240) Conversation logbook (241) Report generation unit (242)

Claims

Claim 1 The system includes a user data collection unit (100) for collecting real-time reading response data of a user, and an interactive reading server (200) for receiving data from the user data collection unit (100), processing it in real-time, and storing and analyzing the user's conversation logs and reading discourse logs. The interactive reading server (200) includes an AI persona generation module (210) for creating an AI conversational subject that mimics the unique writing style, vocabulary, and logical development method of a specific author, a user state analysis module (220) for receiving real-time response data from the user data collection unit (100) and inferring the user's current emotional and cognitive state, a conversation flow control module (230) for dynamically adjusting the content and method of conversation by linking the AI ​​persona generation module (210) and the user state analysis module (220), and a reading activity management module (240) for recording all conversation content between the user and the AI ​​author persona and providing personalized feedback based thereon. The AI ​​persona generation module (210) The system includes a writer data analysis unit (211) for collecting text data including books, interviews, and critiques to database data on the writer's unique writing style, vocabulary, logical structure, and core ideas, and a persona modeling unit (212) for building an AI writer persona model that generates a response in the writer's style to an input user question or situation by fine-tuning a large-scale language model (LLM) based on the data databased through the writer data analysis unit (211); the user state analysis module (220) includes a real-time data processing unit (221) for receiving raw data such as changes in user facial expressions, eye movement, voice tone and speed, and speech rhythm collected from a user data collection unit (100), and an emotion and cognitive state determination unit (222) for classifying the user into one of various states such as immersion, confusion, boredom, or interest by utilizing a pre-trained emotion analysis model.The above conversation flow control module (230) includes a question-and-answer generation unit (231) for generating questions or explanations with an author style applied by reflecting the content currently being read and the characteristics of the AI ​​author persona model of the AI ​​persona generation module (210), and an interaction adjustment unit (232) that performs the function of generating control signals to adjust in real time the difficulty of the questions, the level of detail of the explanations, and the tone of the responses generated by the question-and-answer generation unit (231) based on the results of the emotion and cognitive state analysis received from the user state analysis module, and the above reading activity management module (240) includes a reading perspective movement tracking module (243), and the above reading perspective movement tracking module (243) constructs a knowledge network by setting concepts, people, events, and ideas identified through name recognition (NER) and relationship extraction algorithms from the text to be read as nodes (V) and edges (E), respectively, maps core keywords extracted from the conversation log to corresponding nodes within the knowledge network, and calculates the shortest path length within the knowledge network for node pairs between consecutive utterances. A writer persona-based emotion-linked reading discourse system characterized by deriving a perspective shift distance value, calculating an indicator including a total shift amount which is the cumulative sum of the perspective shift distance values, the number of times a central node is passed, and the same node repetition ratio, applying equal weights to the total shift amount and the central node pass rate and applying a weight of 1.2 times to the same node repetition ratio, and classifying the user's reading state into thought expansion mode, intensive exploration mode, or balanced exploration mode based on a comprehensive indicator. Claim 2 delete Claim 3 delete Claim 4 delete Claim 5 delete Claim 6 delete

Images