01μ΄λ‘Abstract
π ν΅μ¬ μμ½ (ν λ¬Έλ¨)
π One-Paragraph Summary
λ³Έ λ°λͺ
μ μ¬μ©μκ° μ
λ ₯ν μμ°μ΄ ν
μ€νΈ(μ±ν
λ©μμ§, μΌκΈ°, μκ°λ³΄κ³ )μμ μΆμΆν 17κ° μΈμ΄ νΉμ§(νκ· μ΄ν 볡μ‘λ, λ¬Έμ₯ κΈΈμ΄ λΆμ°, μΆμ λͺ
μ¬ λΉμ¨, λ©νμΈμ§ λ§μ»€ λΉλ, μΈκ³Ό μ°κ²°μ¬ λ°λ λ±)μ μ
λ ₯μΌλ‘ νμ¬, Harvard λν Kurt Fischerμ λμ κΈ°μ μ΄λ‘ (Dynamic Skill Theory)μ κΈ°λ°ν 13λ¨κ³(Tier 0β12) λ°λ¬ μμ€μ κ°μ€ ν© + μκ·Έλͺ¨μ΄λ νμ±ν ν¨μλ₯Ό ν΅ν΄ μΆμ νκ³ , μΆμ λ μμ€μ λ°λΌ AI μ±λ΄ μμ€ν
μ μλ΅ μμ± μ (a) μ΄ν 볡μ‘λ, (b) λ¬Έμ₯ ꡬ쑰 κΉμ΄, (c) λΉμ μ μΆμ μ¬μ© λΉμ¨, (d) λ©νμΈμ§ μ λ λΉλλ₯Ό μλ μ‘°μ νμ¬ μ¬μ©μμ λ°λ¬ μμ(Zone of Proximal Development, +1 tier)μμμ νμ΅Β·μΉλ£ ν¨κ³Όλ₯Ό κ·Ήλννλ μ»΄ν¨ν° ꡬν λ°©λ² λ° μμ€ν
μ κ΄ν κ²μ΄λ€.
The present invention relates to a computer-implemented method and system that extracts seventeen (17) linguistic features from natural language input (chat messages, journal entries, self-reports) submitted by a user β including mean lexical complexity, sentence length variance, abstract noun ratio, metacognitive marker frequency, and causal connective density β and infers a developmental level on a thirteen-tier scale (Tier 0 through Tier 12) derived from Harvard professor Kurt Fischer's Dynamic Skill Theory, via a weighted-sum-plus-sigmoid activation function; based on the inferred tier, the system dynamically calibrates an AI chatbot's response-generation parameters β specifically (a) vocabulary complexity, (b) syntactic structure depth, (c) figurative/abstract usage ratio, and (d) metacognitive prompt frequency β to maintain conversation within the user's Zone of Proximal Development (tier + 1), thereby maximizing therapeutic and educational effect.
02λ°λͺ
λΆμΌField of Invention
λ³Έ λ°λͺ
μ μΈκ³΅μ§λ₯ κΈ°λ° λμ§νΈ μ μ κ±΄κ° μλΉμ€(Digital Mental Health) λ° μ μν νμ΅ μμ€ν
(Adaptive Learning Systems) λΆμΌμ μνλ€. λμ± κ΅¬μ²΄μ μΌλ‘λ, μ¬μ©μμ μΈμ΄ ννμΌλ‘λΆν° λ°λ¬ μμ€μ μΆμ νκ³ , κ·Έ μμ€μ λ§μΆ° AI λν μμ΄μ νΈ(LLM κΈ°λ° μ±λ΄)μ μλ΅ λ§€κ°λ³μλ₯Ό μλ μ‘°μ νλ μμ€ν
μ κ΄ν κ²μ΄λ€.
The present invention pertains to the fields of AI-based Digital Mental Health Services and Adaptive Learning Systems. More specifically, it relates to a system that infers a user's developmental level from linguistic expression and automatically adjusts the response parameters of an AI conversational agent (an LLM-based chatbot) to match said level.
κ΄λ ¨ κΈ°μ λΆμΌRelated Technical Fields
- NLP (Natural Language Processing): μΈμ΄ νΉμ§ μΆμΆ, ν
μ€νΈ 볡μ‘λ λΆμ, μλ―Έ λ¨μ νμ±linguistic feature extraction, textual complexity analysis, semantic unit parsing
- Developmental Psychology: Fischer λμ κΈ°μ μ΄λ‘ , Piaget μΈμ§ λ°λ¬ λ¨κ³, Vygotskyμ κ·Όμ λ°λ¬ μμ(ZPD)Fischer's Dynamic Skill Theory, Piaget's stages of cognitive development, Vygotsky's Zone of Proximal Development (ZPD)
- Conversational AI: λν μΈμ΄ λͺ¨λΈ(LLM), ν둬ννΈ μμ§λμ΄λ§, μλ΅ λ§€κ°λ³μνLarge Language Models (LLMs), prompt engineering, response parameterization
- Clinical Psychology / Education: μΉλ£μ λλ§Ή, λ°λ¬μ μ ν©μ±, μΈμ§νλμΉλ£(CBT) λͺ¨λνtherapeutic alliance, developmental appropriateness, cognitive-behavioral therapy (CBT) modularization
03λ°°κ²½ κΈ°μ Background
3.1 Fischer λμ κΈ°μ μ΄λ‘ 3.1 Fischer's Dynamic Skill Theory
Kurt W. Fischer κ΅μ(Harvard Graduate School of Education)λ 1980λ
"A theory of cognitive development: The control and construction of hierarchies of skills"(Psychological Review, 87(6), 477β531)μμ μΈκ°μ μΈμ§Β·μ μ λ°λ¬μ΄ 13λ¨κ³μ μκ³μ κΈ°μ ꡬ쑰λ₯Ό λ°λ₯΄λ©°, κ° λ¨κ³λ λΉμ νμ Β·λ§₯λ½ μμ‘΄μ μΌλ‘ λ°λ¬νλ€λ μ΄λ‘ μ μ μνλ€.
Professor Kurt W. Fischer (Harvard Graduate School of Education) proposed in his seminal 1980 paper, "A theory of cognitive development: The control and construction of hierarchies of skills" (Psychological Review, 87(6), 477β531), that human cognitive-affective development follows a hierarchical skill structure across thirteen tiers, with each tier developing nonlinearly and context-dependently.
| κ³μΈ΅Tier |
λ¨κ³λͺ
Stage Name |
μ νμ μ°λ ΉTypical Age |
μΈμ΄μ νΉμ§Linguistic Marker |
| 0 | Single Reflexes | 0β4mo | μΈμ΄ μ΄μ pre-verbal |
| 1 | Reflex Mappings | 4β8mo | μΈμ΄ μ΄μ pre-verbal |
| 2 | Reflex Systems | 8β12mo | λ¨μ΄ λ¨νΈword fragments |
| 3 | Single Sensorimotor Actions | 1β2y | λ¨μΌ λ¨μ΄single words |
| 4 | Sensorimotor Mappings | 2y | 2-3 λ¨μ΄ κ²°ν©two- to three-word combinations |
| 5 | Sensorimotor Systems | 3β4y | λ¨μ λ¬Έμ₯simple sentences |
| 6 | Single Representations | 4β6y | ꡬ체μ λͺ
μ¬Β·λμ¬concrete nouns and verbs |
| 7 | Representational Mappings | 6β10y | κ΄κ³Β·μΈκ³Ό μ°κ²°relational and causal connectives |
| 8 | Representational Systems | 10β12y | 쑰건·κ°μ€ ννconditional and hypothetical expressions |
| 9 | Single Abstractions | 12β15y | μΆμ λͺ
μ¬abstract nouns |
| 10 | Abstract Mappings | 15β19y | λ©νμΈμ§ ννmetacognitive expressions |
| 11 | Abstract Systems | 19β24y | λ€κ΄μ ν΅ν©multi-perspective integration |
| 12 | Single Principles | 24y+ | μλ¦¬Β·μ²΄κ³ ν΅ν©principle and system synthesis |
Fischer μ΄λ‘ μ ν΅μ¬ ν΅μ°°μ λ°λ¬μ΄ "λ¨μΌ μμμ μΌμ§μ μ§ν"μ΄ μλλΌ "λ§₯λ½λ³Β·κΈ°λ₯λ³λ‘ λ€λ₯Έ μμ€μμ λμ μλ"νλ€λ μ μ΄λ€. λ°λΌμ λμΌν μ¬μ©μλ μ μμ ννμμλ Tier 8, μΆμ μ¬κ³ μμλ Tier 11μ μμΉν μ μλ€.
The core insight of Fischer's theory is that development does not proceed as a "linear advance through a single domain"; rather, it operates simultaneously at different levels across contexts and functions. Hence the same user may operate at Tier 8 for emotional expression while operating at Tier 11 for abstract reasoning.
3.2 ZPD (κ·Όμ λ°λ¬ μμ) μ리3.2 ZPD (Zone of Proximal Development) Principle
Vygotsky(1978)μ ZPD κ°λ
μ νμ΅ ν¨κ³Όκ° κ·Ήλνλλ μ§μ μ΄ νμ¬ μμ€λ³΄λ€ μ½κ° λμ(νμ¬ + 1) μμμμ μμ¬νλ€. λ무 μ¬μ°λ©΄ μ§λ£¨νκ³ νμ΅μ΄ μΌμ΄λμ§ μμΌλ©°, λ무 μ΄λ €μ°λ©΄ μ’μ κ³Ό ννΌκ° λ°μνλ€.
Vygotsky's ZPD concept (1978) holds that learning is maximized when material is positioned slightly above the current level (current + 1). Material too easy yields boredom and no learning; material too difficult yields frustration and avoidance.
3.3 κΈ°μ‘΄ LLM μ±λ΄μ νκ³3.3 Limitations of Existing LLM Chatbots
νμ¬ μ¬μ© μ€μΈ GPT-4, Claude, Gemini λ±μ LLM μ±λ΄μ λͺ¨λ μ¬μ©μμ λ°λ¬ μμ€μ 무μνκ³ κ³ μ λ μΆλ ₯ μ€νμΌλ‘ μλ΅νλ€. μ μ 건κ°Β·κ΅μ‘ λλ©μΈμμ μ΄λ λ€μ λ¬Έμ λ₯Ό λ°μμν¨λ€:
Currently deployed LLM chatbots (GPT-4, Claude, Gemini, and similar systems) respond in a fixed output style that disregards the user's developmental level. In mental health and educational domains, this produces the following problems:
- μΉλ£ λλ§Ή μ½ν: 13μΈ μ²μλ
μκ² Tier 11β12 μ΄νλ‘ λ΅νλ©΄ μνκ°Β·μμΈκ°μ μ λ°νλ€.Weakened therapeutic alliance: answering an adolescent at Tier 11β12 vocabulary triggers feelings of threat and alienation.
- νμ΅ λΉν¨μ¨: λ°μ¬κΈ μ¬μ©μμκ² Tier 6 μμ€ μλ΅μ 무μ±μΒ·λΉμ λ¬Έμ μΌλ‘ μΈμλλ€.Inefficient learning: Tier 6 responses to a doctoral-level user are perceived as superficial and unprofessional.
- μΉλ£ λλ§Ή μ νΈ μμ€: μλ΅ λ§€κ°λ³μκ° μ ν©λμ§ μμΌλ©΄ μ¬μ©μκ° "μ΄ν΄λ°μ§ λͺ»νλ€"λ κ°κ°μ λ°λλ€.Loss of alliance signal: mismatched response parameters give the user a felt sense of "not being understood."
κΈ°μ‘΄ LLM μμ€ν
μ "system prompt"λ₯Ό μλμΌλ‘ μ€μ νμ¬ ν€μ μ‘°μ ν μ μμ§λ§, μ΄λ (a) μ¬μ©μ μ
λ ₯μΌλ‘λΆν° μλ μΆμ λμ§ μμΌλ©°, (b) λν μ§ν μ€ λμ μΌλ‘ κ°±μ λμ§ μμΌλ©°, (c) Fischer 13λ¨κ³μ κ°μ λ°λ¬ μ΄λ‘ μ κ·Όκ±°κ° μλ€.
Existing LLM systems allow manual tone adjustment via a "system prompt," but such tuning (a) is not automatically inferred from user input, (b) is not dynamically updated as the conversation progresses, and (c) is not grounded in any developmental theoretical framework comparable to Fischer's thirteen tiers.
04ν΄κ²° κ³Όμ Problem Statement
β οΈ λ³Έ λ°λͺ
μ΄ ν΄κ²°νλ €λ ν΅μ¬ λ¬Έμ β οΈ Core Problems Addressed by the Invention
-
μλ λ°λ¬ μμ€ μΆμ μ λΆμ¬.Absence of automatic developmental-level inference.
νμ¬ μ΄λ€ LLM μμ€ν
λ μ¬μ©μμ λ°λ¬ μμ€μ μλ μΆμ νμ¬ μλ΅ λ§€κ°λ³μλ₯Ό μ‘°μ νμ§ μλλ€.
No existing LLM system automatically infers a user's developmental level and adjusts response parameters accordingly.
-
νμ€νλ 13λ¨κ³ λ§€νμ λΆμ¬.Absence of a standardized 13-tier mapping.
μΈμ΄ νΉμ§ β λ°λ¬ λ¨κ³ λ§€νμ μ 곡νλ μκ³ λ¦¬μ¦μ΄ μμ
μ μΌλ‘ μ‘΄μ¬νμ§ μλλ€.
No commercially available algorithm provides a linguistic-feature-to-developmental-tier mapping.
-
λ§₯λ½λ³ μ°¨λ± μ μμ λΆμ¬.Absence of context-specific differential adaptation.
μ¬μ©μκ° μ μΒ·μΈμ§Β·κ΄κ³ μμλ§λ€ λ€λ₯Έ λ°λ¬ μμ€μ μμ μ μμμ λ°μνλ μμ€ν
μ΄ λΆμ¬νλ€.
No system reflects the fact that a user may operate at different developmental tiers across emotional, cognitive, and relational contexts.
-
μ€μκ° λμ κ°±μ μ λΆμ¬.Absence of real-time dynamic updating.
λνκ° μ§νλ¨μ λ°λΌ μ¬μ©μμ νν μμ€μ΄ λ³νν μ μμΌλ, μ΄λ₯Ό μ¦μ λ°μνλ λ©μ»€λμ¦μ΄ μλ€.
A user's expressive level may shift as the conversation progresses, yet no mechanism exists to immediately reflect such shifts.
-
ZPD κΈ°λ° μλ΅ λ³΄μ μ λΆμ¬.Absence of ZPD-based response calibration.
νμ¬ μμ€ + 1λ‘ μλ΅μ μλ 보μ νμ¬ νμ΅Β·μΉλ£ ν¨κ³Όλ₯Ό κ·Ήλννλ μμ€ν
μ΄ λΆμ¬νλ€.
No system performs ZPD-based calibration β automatically setting the response one tier above the user's current level β to maximize learning and therapeutic effect.
05λ°λͺ
μμ½Summary of Invention
π λ°λͺ
μ ν΅μ¬ ꡬμ±π Core Components of the Invention
5.1 μμ€ν
κ΅¬μ± μμ5.1 System Components
- μΈμ΄ νΉμ§ μΆμΆ λͺ¨λ (LFE)Linguistic Feature Extractor (LFE): μ¬μ©μ μ
λ ₯ ν
μ€νΈλ‘λΆν° 17κ° μ λ νΉμ§μ μΆμΆextracts seventeen (17) quantitative features from user input text
- Tier μΆλ‘ μμ§ (TIE)Tier Inference Engine (TIE): κ°μ€ ν© + μκ·Έλͺ¨μ΄λλ₯Ό ν΅ν΄ 0β12 λ¨κ³λ₯Ό μΆμ infers a 0β12 tier via weighted-sum-plus-sigmoid activation
- λ§₯λ½ λΆλ¦¬ λͺ¨λ (CSM)Context Separation Module (CSM): μ μΒ·μΈμ§Β·κ΄κ³ λλ©μΈλ³λ‘ λ³λ Tier μΆμ estimates separate Tiers for emotional, cognitive, and relational domains
- μλ΅ λ§€κ°λ³μ λ§€νΌ (RPM)Response Parameter Mapper (RPM): Tier β AI μλ΅ λ§€κ°λ³μ(μ΄ν 볡μ‘λ, λ¬Έμ₯ κΉμ΄, μΆμ λΉμ¨, λ©νμΈμ§ λΉλ) λ§€νmaps Tier values to AI response parameters (vocabulary complexity, syntactic depth, abstraction ratio, metacognitive prompt frequency)
- ZPD 보μ κΈ° (ZAC)ZPD Adjuster (ZAC): μλ΅ Tier = μ¬μ©μ Tier + 1λ‘ λ³΄μ calibrates response Tier as user Tier + 1
- λμ κ°±μ 루ν (DUL)Dynamic Update Loop (DUL): λ§€ λ©μμ§λ§λ€ Tier μ¬μΆμ λ° μλ΅ λ§€κ°λ³μ κ°±μ re-estimates Tier and updates response parameters on each message
5.2 ν΅μ¬ μ°¨λ³μ (Inventive Step)5.2 Inventive Steps
- λ°λ¬ μ΄λ‘ κ³Ό LLM κ²°ν©:Coupling developmental theory with LLMs: Fischer 13λ¨κ³ + LLM μλ΅ λ§€κ°λ³μμ μ§μ λ§€νμ μ ν κΈ°μ μ μ‘΄μ¬νμ§ μλλ€.A direct mapping between Fischer's thirteen tiers and LLM response parameters does not exist in the prior art.
- λ§₯λ½ λΆλ¦¬ μΆμ :Context-separated inference: λ¨μΌ μ μκ° μλ λλ©μΈλ³ λ€μ°¨μ Tier 벑ν°.a multi-dimensional Tier vector across domains, not a single score.
- μ€μκ° ZPD 보μ :Real-time ZPD calibration: λ§€ μλ΅λ§λ€ +1 보μ μΌλ‘ νμ΅ μμ μ μ§.+1 calibration applied at each response to keep dialogue within the learning zone.
- μΈμ΄ νΉμ§ κ°μ€μΉμ μμμ κ²μ¦:Clinically validated linguistic-feature weights: Boston Neuromind μμ λ°μ΄ν°λ‘ 보μ λ 17κ° νΉμ§ κ°μ€μΉ.weights for the seventeen features calibrated against Boston Neuromind clinical data.
06μμΈ μ€λͺ
Detailed Description
6.1 17κ° μΈμ΄ νΉμ§μ μ μ6.1 Definition of the Seventeen Linguistic Features
κ° μ¬μ©μ μ
λ ₯ λ©μμ§λ‘λΆν° λ€μ 17κ° μ λ νΉμ§μ μΆμΆνλ€:From each user input message, the following seventeen quantitative features are extracted:
| # |
νΉμ§λͺ
Feature |
μ μDefinition |
λ°λ¬ μκ·ΈλDevelopmental Signal |
| F1 | MLD (Mean Lexical Difficulty) | λ¨μ΄λ³ λΉλ μμ νκ· average inverse word-frequency | Tier β |
| F2 | SLV (Sentence Length Variance) | λ¬Έμ₯ κΈΈμ΄ λΆμ°variance in sentence length | Tier β |
| F3 | ANR (Abstract Noun Ratio) | μ 체 λͺ
μ¬ μ€ μΆμλͺ
μ¬ λΉμ¨ratio of abstract nouns among all nouns | Tier 9+ |
| F4 | MMF (Metacognitive Marker Frequency) | "λ΄ μκ°", "λλμ" λ± λΉλfrequency of markers such as "I think," "I feel like" | Tier 10+ |
| F5 | CCD (Causal Connective Density) | "λ°λΌμ", "μλνλ©΄" λ°λdensity of "therefore," "because" connectives | Tier 7+ |
| F6 | CHD (Conditional Hypothetical Density) | "λ§μ½ ~λΌλ©΄" λΉλfrequency of "if-then" constructions | Tier 8+ |
| F7 | SCD (Subordinate Clause Depth) | μ’
μμ νκ· κΉμ΄average depth of subordinate clauses | Tier β |
| F8 | FUR (Figurative Usage Ratio) | μμ Β·λΉμ λΉμ¨ratio of metaphors and similes | Tier 9+ |
| F9 | MPI (Multi-Perspective Indicator) | "ννΈμΌλ‘λ" λ± λ€κ΄μ λ§μ»€multi-perspective markers such as "on the other hand" | Tier 11+ |
| F10 | EVL (Emotion Vocabulary Lexicon) | μ μ μ΄ν νλΆλrichness of emotional vocabulary | Tier 9+ |
| F11 | TPS (Temporal Perspective Shifts) | κ³Όκ±°Β·λ―ΈλΒ·κ°μ μμ μ ν λΉλfrequency of pastβfutureβhypothetical tense shifts | Tier 8+ |
| F12 | NEG (Negation Complexity) | μ€μ²© λΆμ μ¬μ©use of nested negation | Tier 8+ |
| F13 | QDF (Question Depth Factor) | μ§λ¬Έμ μΆμλabstractness of questions posed | Tier β |
| F14 | SLF (Self-Reflection Frequency) | μκΈ° μ°Έμ‘° + νκ° λ§μ»€frequency of self-reference + evaluation markers | Tier 10+ |
| F15 | SCC (System Concept Count) | μ²΄κ³ κ°λ
μ΄νcount of system-level concept terms | Tier 12 |
| F16 | PRC (Principle Reference Count) | μ리 μ°Έμ‘° ννreferences to principles | Tier 12 |
| F17 | NRD (Narrative Reasoning Depth) | μμ¬μ μΈκ³Ό κΉμ΄depth of narrative causal reasoning | Tier β |
6.2 Tier μΆλ‘ μμ6.2 Tier Inference Formula
λ°λ¬ λ¨κ³ Tier T(x)λ λ€μ μμΌλ‘ κ³μ°νλ€:The developmental tier T(x) is computed as follows:
μμ 1: Tier μΆμ ν¨μEquation 1: Tier Inference Function
T(x) = round( 12 Β· Ο( Ξ£α΅’ wα΅’ Β· Fα΅’(x) + b ) )
where:
x = user input text
Fα΅’ = i-th linguistic feature (i = 1..17)
wα΅’ = trained weight for feature i (calibrated on Boston Neuromind clinical corpus)
b = bias term
Ο = logistic sigmoid: Ο(z) = 1 / (1 + e^(-z))
round(Β·) = nearest integer in {0, 1, 2, ..., 12}
6.3 λ§₯λ½ λΆλ¦¬ μΆμ 6.3 Context-Separated Inference
λμΌ μ¬μ©μμ λν΄ 3κ° λλ©μΈλ³λ‘ λ
립μ μΈ Tierλ₯Ό μ°μΆνλ€:Three independent Tiers are computed per user, one for each domain:
T_emotional = T(x, weights_emotional)
T_cognitive = T(x, weights_cognitive)
T_relational = T(x, weights_relational)
β Tier Vector V = (T_e, T_c, T_r)
κ° λλ©μΈλ³ κ°μ€μΉ(weights_emotional, weights_cognitive, weights_relational)λ 17κ° νΉμ§ μ€ ν΄λΉ λλ©μΈμ λ κ°ν μκ·Έλμ μ£Όλ νΉμ§μ κ°μ‘°νλ€.Each domain weight set (weights_emotional, weights_cognitive, weights_relational) emphasizes the features that provide the strongest signal for that domain among the seventeen features.
6.4 μλ΅ λ§€κ°λ³μ λ§€ν6.4 Response Parameter Mapping
μΆμ λ Tierμ λ°λΌ LLM μλ΅ μμ± μ λ€μ 4κ° λ§€κ°λ³μκ° μλ μ‘°μ λλ€:Based on the inferred Tier, the following four parameters are automatically adjusted at the time of LLM response generation:
| λ§€κ°λ³μParameter |
Tier 0β4 |
Tier 5β8 |
Tier 9β12 |
| P1: μ΄ν 볡μ‘λVocabulary Complexity |
κΈ°μ΄ 1000 λ¨μ΄basic 1,000-word |
νμ€ 3000 λ¨μ΄standard 3,000-word |
κ³ κΈ 6000+ λ¨μ΄advanced 6,000+ word |
| P2: λ¬Έμ₯ κΉμ΄Syntactic Depth |
λ¨μλ¬Έ μμ£Όpredominantly simple |
볡문 νμ©complex clauses allowed |
μ€μ²© μ’
μμ nested subordinate clauses |
| P3: μΆμ λΉμ¨Abstraction Ratio |
0β10% |
10β35% |
35β70% |
| P4: λ©νμΈμ§ μ λMetacognitive Prompts |
μμnone |
κ°νμ intermittent |
μμ£Όfrequent |
6.5 ZPD 보μ (+1 Rule)6.5 ZPD Calibration (the "+1" Rule)
T_response = min( T_user + 1, 12 )
β Response is generated using parameter set for T_response,
not T_user, ensuring conversation stays in Zone of Proximal Development
μ: μ¬μ©μ Tier 7 (10μΈ μλ μμ€) β μλ΅ Tier 8 (쑰건·κ°μ€ μμ€ ν λ¨κ³ μΆκ°) μΌλ‘ μμ±νμ¬ μ¬μ©μκ° μ μΆλ‘ ν¨ν΄μ λμ νλλ‘ μ λνλ€.Example: a user at Tier 7 (10-year-old level) is met with a response generated at Tier 8 (introducing one level of conditional/hypothetical reasoning above current capacity), thereby inviting the user to engage with a new reasoning pattern.
6.6 λμ κ°±μ 루ν6.6 Dynamic Update Loop
λ§€ μ¬μ©μ λ©μμ§ μμ μ λ€μ μ μ°¨λ₯Ό λ°λ³΅νλ€:The following procedure is iterated upon receipt of each user message:
- ν
μ€νΈ μ
λ ₯ β 17κ° νΉμ§ μΆμΆ (LFE)Text input β extract seventeen features (LFE)
- 3κ° λλ©μΈλ³ Tier μΆμ (TIE + CSM)Estimate Tier for three domains (TIE + CSM)
- μ΅κ·Ό 5κ° λ©μμ§ μ΄λ νκ· μΌλ‘ Tier μμ νStabilize Tier via moving average over the most recent five messages
- +1 보μ μ μ© (ZAC)Apply +1 calibration (ZAC)
- μλ΅ λ§€κ°λ³μ μ
λ°μ΄νΈ (RPM)Update response parameters (RPM)
- LLMμ λ§€κ°λ³μ μ μ©λ system prompt μ£Όμ
ν μλ΅ μμ±Generate response after injecting parameter-adjusted system prompt into the LLM
08μ²κ΅¬νClaims
μ²κ΅¬ν 1 (λ
립ν)Claim 1 (Independent)
μ¬μ©μμ μμ°μ΄ μ
λ ₯μ μλ΅νλ μΈκ³΅μ§λ₯ λν μμ€ν
μ μν΄ μνλλ, μ»΄ν¨ν° ꡬν μ μν μλ΅ μμ± λ°©λ²μΌλ‘μ:
(a) μ¬μ©μ μ
λ ₯ ν
μ€νΈλ‘λΆν° νκ· μ΄ν 볡μ‘λ, λ¬Έμ₯ κΈΈμ΄ λΆμ°, μΆμ λͺ
μ¬ λΉμ¨, λ©νμΈμ§ λ§μ»€ λΉλ λ° μΈκ³Ό μ°κ²°μ¬ λ°λλ₯Ό ν¬ν¨νλ 볡μμ μ λμ μΈμ΄ νΉμ§μ μΆμΆνλ λ¨κ³;
(b) μΆμΆλ μΈμ΄ νΉμ§μ κ°μ€ ν©μ λΉμ ν νμ±ν ν¨μλ₯Ό μ μ©νμ¬, Fischer λμ κΈ°μ μ΄λ‘ μμ μ μλ 0λΆν° 12κΉμ§μ 13κ° λ°λ¬ λ¨κ³ μ€ μ¬μ©μμ νμ¬ λ°λ¬ λ¨κ³ μΆμ κ°μ μ°μΆνλ λ¨κ³;
(c) μΆμ λ λ°λ¬ λ¨κ³μ 1μ κ°μ°νμ¬ μλ΅ λ¨κ³λ₯Ό κ²°μ νλ, μ΅λκ°μ 12λ‘ μ ννλ λ¨κ³;
(d) κ²°μ λ μλ΅ λ¨κ³μ λ°λΌ μΈκ³΅μ§λ₯ μλ΅ μμ±κΈ°μ (i) μ΄ν 볡μ‘λ, (ii) λ¬Έμ₯ ꡬ쑰 κΉμ΄, (iii) μΆμ νν λΉμ¨ λ° (iv) λ©νμΈμ§ μ λ λΉλλ₯Ό ν¬ν¨νλ μλ΅ λ§€κ°λ³μ μ§ν©μ μλ μ‘°μ νλ λ¨κ³; λ°
(e) μ‘°μ λ μλ΅ λ§€κ°λ³μλ₯Ό μ μ©νμ¬ μλ΅ ν
μ€νΈλ₯Ό μμ±νλ λ¨κ³;
λ₯Ό ν¬ν¨νλ κ²μ νΉμ§μΌλ‘ νλ λ°©λ².
A computer-implemented method, performed by an artificial intelligence conversational system responding to natural language input from a user, the method comprising:
(a) extracting, from a user input text, a plurality of quantitative linguistic features including mean lexical complexity, sentence length variance, abstract noun ratio, metacognitive marker frequency, and causal connective density;
(b) computing a weighted sum of the extracted linguistic features and applying a nonlinear activation function thereto, thereby yielding an estimated current developmental tier of the user, said tier being one of thirteen tiers (Tier 0 through Tier 12) defined in Fischer's Dynamic Skill Theory;
(c) determining a response tier by adding one (1) to the estimated current developmental tier, subject to a maximum of twelve (12);
(d) automatically adjusting a set of response parameters of an artificial intelligence response generator according to the determined response tier, said set comprising (i) vocabulary complexity, (ii) syntactic structure depth, (iii) abstract-expression ratio, and (iv) metacognitive prompt frequency; and
(e) generating a response text by applying the adjusted response parameters.
μ²κ΅¬ν 2 (μ’
μν)Claim 2 (Dependent)
μ²κ΅¬ν 1μ μμ΄μ, λ¨κ³ (a)λ 17κ°μ μ λμ μΈμ΄ νΉμ§μ μΆμΆνκ³ , μκΈ° 17κ° νΉμ§μ νκ· μ΄ν λμ΄λ, λ¬Έμ₯ κΈΈμ΄ λΆμ°, μΆμ λͺ
μ¬ λΉμ¨, λ©νμΈμ§ λ§μ»€ λΉλ, μΈκ³Ό μ°κ²°μ¬ λ°λ, 쑰건/κ°μ€ νν λ°λ, μ’
μμ κΉμ΄, λΉμ μ¬μ© λΉμ¨, λ€κ΄μ μ§ν, μ μ μ΄ν νλΆλ, μμ μ ν λΉλ, λΆμ 볡μ‘λ, μ§λ¬Έ κΉμ΄, μκΈ°μ±μ°° λΉλ, μ²΄κ³ κ°λ
μ, μ리 μ°Έμ‘° μ λ° μμ¬μ μΆλ‘ κΉμ΄λ₯Ό ν¬ν¨νλ κ²μ νΉμ§μΌλ‘ νλ λ°©λ².
The method of Claim 1, wherein step (a) extracts seventeen (17) quantitative linguistic features, said seventeen features comprising: mean lexical difficulty; sentence length variance; abstract noun ratio; metacognitive marker frequency; causal connective density; conditional/hypothetical density; subordinate clause depth; figurative usage ratio; multi-perspective indicator; emotion vocabulary richness; temporal perspective shifts; negation complexity; question depth factor; self-reflection frequency; system concept count; principle reference count; and narrative reasoning depth.
μ²κ΅¬ν 3 (μ’
μν)Claim 3 (Dependent)
μ²κ΅¬ν 1μ μμ΄μ, λ¨κ³ (b)λ μ μμ λλ©μΈ, μΈμ§μ λλ©μΈ λ° κ΄κ³μ λλ©μΈμ λν΄ κ°κ° λ³κ°μ κ°μ€μΉ μ§ν©μ μ μ©νμ¬ 3κ°μ λ
립μ μΈ λ°λ¬ λ¨κ³ μΆμ κ°μ μ°μΆνκ³ , λ¨κ³ (d)μ μλ΅ λ§€κ°λ³μ μ‘°μ μ λλ©μΈλ³ μΆμ κ°μ κΈ°λ°νμ¬ μνλλ κ²μ νΉμ§μΌλ‘ νλ λ°©λ².
The method of Claim 1, wherein step (b) computes three independent developmental-tier estimates by applying distinct weight sets for an emotional domain, a cognitive domain, and a relational domain, respectively, and wherein the response-parameter adjustment of step (d) is performed based on the domain-specific estimates.
μ²κ΅¬ν 4 (μ’
μν)Claim 4 (Dependent)
μ²κ΅¬ν 1μ μμ΄μ, μ¬μ©μμμ λνκ° μ§νλλ λμ λ§€ μ¬μ©μ μ
λ ₯ λ©μμ§λ§λ€ λ¨κ³ (a) λ΄μ§ (e)λ₯Ό λ°λ³΅νκ³ , λ¨κ³ (b)μ λ°λ¬ λ¨κ³ μΆμ κ°μ μ΅κ·Ό Nκ°(Nβ₯2) λ©μμ§μ λν μ΄λ νκ· μΌλ‘ μμ ννλ κ²μ νΉμ§μΌλ‘ νλ λ°©λ².
The method of Claim 1, further comprising iterating steps (a) through (e) for each user input message during the course of a conversation, and stabilizing the developmental-tier estimate of step (b) by computing a moving average over the most recent N messages, where N is greater than or equal to two.
μ²κ΅¬ν 5 (μ’
μν)Claim 5 (Dependent)
μ²κ΅¬ν 1μ μμ΄μ, λ¨κ³ (b)μ λΉμ ν νμ±ν ν¨μλ λ‘μ§μ€ν± μκ·Έλͺ¨μ΄λ ν¨μμΈ Ο(z) = 1 / (1 + e^(-z)) μ΄κ³ , λ°λ¬ λ¨κ³ μΆμ κ°μ 12 Β· Ο(κ°μ€ν© + νΈν₯)μ κ°μ₯ κ°κΉμ΄ μ μλ‘ λ°μ¬λ¦Όνμ¬ κ²°μ λλ κ²μ νΉμ§μΌλ‘ νλ λ°©λ².
The method of Claim 1, wherein the nonlinear activation function of step (b) is the logistic sigmoid function Ο(z) = 1 / (1 + e^(-z)), and the developmental-tier estimate is determined by rounding 12Β·Ο(weighted sum + bias) to the nearest integer.
μ²κ΅¬ν 6 (μ’
μν)Claim 6 (Dependent)
μ²κ΅¬ν 1μ μμ΄μ, λ¨κ³ (e)μ μλ΅ ν
μ€νΈ μμ±μ λν μΈμ΄ λͺ¨λΈ(LLM)μ μμ€ν
ν둬ννΈλ₯Ό μ£Όμ
ν¨μΌλ‘μ¨ μνλλ©°, μκΈ° μμ€ν
ν둬ννΈλ λ¨κ³ (d)μμ κ²°μ λ μλ΅ λ§€κ°λ³μμ λ°λΌ λμ μΌλ‘ ꡬμ±λλ κ²μ νΉμ§μΌλ‘ νλ λ°©λ².
The method of Claim 1, wherein the response text generation of step (e) is performed by injecting a system prompt into a Large Language Model (LLM), said system prompt being dynamically composed according to the response parameters determined in step (d).
μ²κ΅¬ν 7 (λ
립ν β μμ€ν
)Claim 7 (Independent β System)
μ²κ΅¬ν 1 λ΄μ§ 6 μ€ μ΄λ ν νμ λ°©λ²μ μννκΈ° μν, μ μ΄λ νλμ νλ‘μΈμμ μκΈ° νλ‘μΈμμ μν΄ μ€νλλ λͺ
λ Ήμ΄λ₯Ό μ μ₯νλ λΉμΌμμ μ»΄ν¨ν° νλ
κ°λ₯ μ μ₯ 맀체λ₯Ό ν¬ν¨νλ μ μν μΈκ³΅μ§λ₯ λν μμ€ν
.
An adaptive artificial intelligence conversational system for performing the method of any one of Claims 1 through 6, the system comprising at least one processor and a non-transitory computer-readable storage medium storing instructions executable by the processor.
μ²κ΅¬ν 8 (λ
립ν β 맀체)Claim 8 (Independent β Medium)
μ»΄ν¨ν°μ μν΄ μ€νλ λ μ²κ΅¬ν 1 λ΄μ§ 6 μ€ μ΄λ ν νμ λ°©λ²μ μννλλ‘ νλ λͺ
λ Ήμ΄λ₯Ό μ μ₯νλ λΉμΌμμ μ»΄ν¨ν° νλ
κ°λ₯ μ μ₯ 맀체.
A non-transitory computer-readable storage medium storing instructions which, when executed by a computer, cause the computer to perform the method of any one of Claims 1 through 6.
09μ ν κΈ°μ λΉκ΅Prior Art Comparison
| μ ν κΈ°μ Prior Art |
μ κ·Ό λ°©μApproach |
νκ³Limitation |
λ³Έ λ°λͺ
κ³Όμ μ°¨μ΄Distinction from Invention |
OpenAI GPT-4 (2023) System prompts |
μλ ν€ μ‘°μ manual tone adjustment |
μλ μΆμ μμ, λ°λ¬ μ΄λ‘ λΆμ¬no auto-inference; no developmental grounding |
17 νΉμ§ μλ μΆμΆ + 13λ¨κ³ λ§€ν17-feature auto-extraction + 13-tier mapping |
| Khan Academy Khanmigo (2023) |
νλ
λ³ μ¬μ μ€μ grade-level presets |
μ μ , μ¬μ©μ μ
λ ₯ κΈ°λ° μ‘°μ μμstatic; no input-based adjustment |
μ€μκ° λμ κ°±μ + ZPD +1real-time dynamic updating + ZPD +1 |
| Replika / Character.AI |
κ³ μ νλ₯΄μλfixed personas |
μ¬μ©μ λ°λ¬ μμ€ λ¬΄κ΄decoupled from user developmental level |
λ§₯λ½λ³ λλ©μΈ λΆλ¦¬ μΆμ domain-separated context-specific inference |
Woebot / Wysa (CBT μ±λ΄) |
κ³ μ μ€ν¬λ¦½νΈfixed scripts |
μΈμ΄ μ μ μμno linguistic adaptation |
κ°λ³ λ©μμ§λ§λ€ λ§€κ°λ³μ κ°±μ per-message parameter updating |
Lexile / Flesch-Kincaid (κ°λ
μ± μ μ) |
ν
μ€νΈ λμ΄λ μΈ‘μ text difficulty scoring |
μΆλ ₯ μΈ‘μ λ§ κ°λ₯, μ
λ ₯βμΆλ ₯ λ§€ν μμoutput measurement only; no input-to-output mapping |
μ¬μ©μ μ
λ ₯ β AI μλ΅ λ§€κ°λ³μ μ§μ λ§€νdirect mapping from user input to AI response parameters |
π― λ°λͺ
μ μ§λ³΄μ± (Inventive Step)π― Inventive Step
λ³Έ λ°λͺ
μ (1) Fischer 13λ¨κ³ λ°λ¬ μ΄λ‘ μ LLM μλ΅ λ§€κ°λ³μμ μ§μ λ§€νν μ΅μ΄μ μμ€ν
μ΄λ©°, (2) μ¬μ©μ μ
λ ₯μΌλ‘λΆν° μλ μΆμ λ λ°λ¬ λ¨κ³μ λ°λΌ μλ΅μ ZPD(+1)λ‘ λ³΄μ νκ³ , (3) μ μΒ·μΈμ§Β·κ΄κ³ λλ©μΈμ λΆλ¦¬νμ¬ λ€μ°¨μ μ μμ μννλ€. μ΄ μΈ κ°μ§ κ²°ν©μ μ ν κΈ°μ μ μ‘΄μ¬νμ§ μμΌλ©°, λ°λͺ
μ λΉμλͺ
μ±(non-obviousness)μ ꡬμ±νλ€.
The present invention is (1) the first system to directly map Fischer's 13-tier developmental theory to LLM response parameters; (2) it calibrates responses to ZPD (+1) based on developmental tiers automatically inferred from user input; and (3) it performs multi-dimensional adaptation by separating the emotional, cognitive, and relational domains. The combination of these three elements does not exist in the prior art and establishes the non-obviousness of the invention.
10μ°μ
μ μ΄μ© κ°λ₯μ±Industrial Applicability
10.1 μ μ© μμ₯10.1 Target Markets
- λμ§νΈ μ μ 건κ°:Digital mental health: μΉλ£ μ±λ΄, μκ° λμ μ±, νλ μ½μΉ (Woebot, Wysa, Replika λ± μΆμ)therapeutic chatbots, self-help apps, behavioral coaching (overtaking Woebot, Wysa, Replika)
- κ΅μ‘ EdTech:Educational EdTech: κ°μΈν νν°λ§, μ μν νμ΅ νλ«νΌ (Khan Academy, Duolingo)personalized tutoring, adaptive learning platforms (Khan Academy, Duolingo)
- μ§μ₯ μ½μΉ:Workplace coaching: 리λμ κ°λ°, μΈμ¬ μ½μΉ μ±λ΄leadership development, HR coaching chatbots
- κ³ κ° μλΉμ€:Customer service: κ³ κ° λ°λ¬ μμ€μ λ§μΆ μλ (νΉν μλμ΄ μΌμ΄)level-matched customer service, especially in senior care
- μμ λꡬ:Clinical tools: μΉλ£μ¬ 보쑰, νμ μκ°λ³΄κ³ λΆμtherapist assistance, patient self-report analysis
10.2 ꡬν νν10.2 Implementations
- SaaS API: λ€λ₯Έ μ±λ΄ νμ¬κ° Tier μΆμ Β·λ§€ν κΈ°λ₯μ APIλ‘ νΈμΆSaaS API: other chatbot vendors invoke tier inference and mapping via API
- B2C: TalkCatcher.com β μ§μ μ μ©ν μ»΄ν¨λμΈ λ΄B2C: TalkCatcher.com β companion bot with direct application
- B2B λΌμ΄μ μ€: μλ£κΈ°κ΄Β·κ΅μ‘κΈ°κ΄ SDK μ 곡B2B licensing: SDK delivery to medical and educational institutions
10.3 μ΄λ‘ μ κΈ°λ°10.3 Theoretical Basis
λ³Έ λ°λͺ
μ λ°λͺ
μ(Boston Neuromind μ€λ¦½μ)κ° Harvard Graduate School of Educationμ Visiting Scholarλ‘μ ζ
Kurt Fischer κ΅μμ μ§μ μ§λνμ λμ κΈ°μ μ΄λ‘ μ νμ΅νκ³ , λν ERP κΈ°λ° νμ΅ μ΅μ ν μ°κ΅¬μ μ°Έμ¬ν κ²½νμ ν λλ‘ κ΅¬μλμλ€. μ΄λ‘ μ κΈ°λ°μ κΉμ΄μ μμμ κ²μ¦μ κ²°ν©μ΄ λ³Έ λ°λͺ
μ λ
μ°½μ±μ λ·λ°μΉ¨νλ€.
The invention was conceived by the inventor β founder of Boston Neuromind β based on direct mentorship with the late Professor Kurt Fischer at the Harvard Graduate School of Education as Visiting Scholar, where the inventor studied Dynamic Skill Theory firsthand and participated in EEG/ERP-based learning-optimization research. The depth of theoretical grounding combined with clinical validation underpins the originality of this invention.
11κ΄λ ¨ λ
Όλ¬ΈReferences
λ³Έ λ°λͺ
μ μ΄λ‘ μ Β·μμμ κ·Όκ±°κ° λλ ν΅μ¬ λ
Όλ¬Έ λ° μλ£. ν΄λ¦νλ©΄ μΈλΆ μΆμ²λ‘ μ΄λν©λλ€.
Key papers and resources providing the theoretical and clinical basis for this invention. Click links to access external sources.
A. Fischer λμ κΈ°μ μ΄λ‘ A. Fischer Dynamic Skill Theory
- A1Fischer KW. A theory of cognitive development: The control and construction of hierarchies of skills. Psychological Review, 1980; 87(6):477-531. DOI ↗
- A2Fischer KW, Bidell TR. Dynamic development of action and thought. In: Damon W, Lerner RM (Eds.), Handbook of Child Psychology, 6th ed. Wiley, 2006; 1:313-399.
- A3Fischer KW, Daley SG. Connecting cognitive science and neuroscience to education: Potentials and pitfalls in inferring executive processes. In: Mind, Brain, and Education in Reading Disorders, Cambridge University Press, 2007; 55-72.
- A4Fischer KW. Dynamic cycles of cognitive and brain development: Measuring growth in mind, brain, and education. In: Battro AM et al. (Eds.), The Educated Brain, Cambridge University Press, 2008; 127-150.
B. Vygotsky ZPD λ° λ°λ¬ νμ΅ μ΄λ‘ B. Vygotsky ZPD and Developmental Learning
- B1Vygotsky LS. Mind in Society: The Development of Higher Psychological Processes. Harvard University Press, 1978.
- B2Wood D, Bruner JS, Ross G. The role of tutoring in problem solving. Journal of Child Psychology and Psychiatry, 1976; 17(2):89-100. DOI ↗
- B3Chaiklin S. The zone of proximal development in Vygotsky's analysis of learning and instruction. In: Kozulin A et al. (Eds.), Vygotsky's Educational Theory in Cultural Context, Cambridge University Press, 2003; 39-64.
C. μΈμ΄ 볡μ‘λ λ° κ°λ
μ± μΈ‘μ C. Linguistic Complexity and Readability
- C1Flesch R. A new readability yardstick. Journal of Applied Psychology, 1948; 32(3):221-233. DOI ↗
- C2McCarthy PM, Jarvis S. MTLD, vocd-D, and HD-D: A validation study of sophisticated approaches to lexical diversity assessment. Behavior Research Methods, 2010; 42(2):381-392. DOI ↗
- C3Lu X. Automatic analysis of syntactic complexity in second language writing. International Journal of Corpus Linguistics, 2010; 15(4):474-496. DOI ↗
- C4Graesser AC, McNamara DS, Louwerse MM, Cai Z. Coh-Metrix: Analysis of text on cohesion and language. Behavior Research Methods, Instruments, & Computers, 2004; 36(2):193-202. DOI ↗
D. LLM λ° λνν AID. LLMs and Conversational AI
- D1Brown TB, Mann B, Ryder N, et al. Language Models are Few-Shot Learners (GPT-3 paper). NeurIPS 2020. arXiv ↗
- D2Devlin J, Chang MW, Lee K, Toutanova K. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. NAACL 2019. arXiv ↗
- D3Bommasani R, Hudson DA, Adeli E, et al. On the Opportunities and Risks of Foundation Models. arXiv:2108.07258, 2021. arXiv ↗
E. μ μν νμ΅ μμ€ν
E. Adaptive Learning Systems
- E1VanLehn K. The Behavior of Tutoring Systems. International Journal of Artificial Intelligence in Education, 2006; 16(3):227-265.
- E2Aleven V, Sewall J. Hands-on tutorial: Constructing tutors with the Cognitive Tutor Authoring Tools. Proceedings of the 2010 ACM CHI Workshop, 2010.
F. λμ§νΈ μ μ 건κ°F. Digital Mental Health
- F1Fitzpatrick KK, Darcy A, Vierhile M. Delivering Cognitive Behavior Therapy to Young Adults with Symptoms of Depression and Anxiety Using a Fully Automated Conversational Agent (Woebot): A Randomized Controlled Trial. JMIR Mental Health, 2017; 4(2):e19. DOI ↗
- F2Inkster B, Sarda S, Subramanian V. An Empathy-Driven, Conversational Artificial Intelligence Agent (Wysa) for Digital Mental Well-Being. JMIR mHealth and uHealth, 2018; 6(11):e12106. DOI ↗