AI 기반 한국어 맞춤형 복습 시스템의 개발 및 적용 -대학 부설 한국어교육기관 내 정규 과정을 중심으로- = Development and Application of an AI-Based Personalized Korean Language Review System ? Focusing on Regular Courses in University-Affiliated Korean Language Institutes ?

다국어 초록 (Multilingual Abstract)

This study designed and developed an AI-based personalized review tool for beginner-level learners at a university-affiliated Korean language education institution, and verified its educational effectiveness and field applicability through implementation in the regular curriculum. The research originated from the recognition that despite the rapid increase in demand for Korean language learning—with the number of international students in Korea exceeding 260,000 in 2024 and TOPIK overseas applicants increasing approximately 4.4 times from 2020 to 2024—the digital transformation of Korean language education remains slow. Particularly in beginner classrooms at university Korean language institutes, structural limitations exist where learners from diverse nationalities, learning backgrounds, and motivations are mixed in a single classroom, making it difficult to provide individualized review opportunities. Additionally, instructors face realistic constraints in confirming individual learners' acquisition due to heavy workloads including class preparation, progress management, assessment, and individual counseling.
The core research questions of this study were as follows. First, what principles should guide the design of an AI tool supporting personalized review for beginner learners? Second, what effects does the developed tool have on learners' learning experiences and academic achievement in actual educational settings? Third, what support does this tool provide to instructors, and what are their perceptions regarding human-AI collaboration? Fourth, based on these cases, what sustainable AI utilization models and practical recommendations can be proposed for Korean language education settings?
To answer these questions, this study reviewed established educational theories including Self-Regulated Learning (SRL), Self-Determination Theory (SDT), spacing effect, and knowledge tracing algorithms, deriving seven core design principles. The first principle was implementing digital scaffolding to support the SRL cycle, specifically supporting Zimmerman's forethought, performance, and self- reflection learning stages in a digital environment. The second principle was designing motivation based on SDT, satisfying learners' needs for autonomy, competence, and relatedness while introducing gamification elements focused on cooperative social comparison and individual growth rather than ranking competition. The third principle was applying cognitive load management and adaptive distributed learning, segmenting learning units to approximately 10-15 minutes following microlearning principles while implementing adaptive distributed learning that calculates optimal review schedules based on individual memory retention states. The fourth principle was applying hybrid algorithms optimized for knowledge types, differentially applying the HLR algorithm specialized in modeling forgetting for vocabulary as declarative knowledge and the BKT algorithm suitable for mastery tracking for grammar as procedural knowledge. The fifth principle was building a human-AI collaboration interface for instructor capability enhancement, implementing an instructor dashboard as a core feature enabling instructors to monitor class status, analyze individual learner data, and review and modify AI feedback. The sixth principle was ensuring close linkage with regular curriculum and accessibility, technically synchronizing the system with the institution's regular textbooks, class progress, and assessment schedules while developing it as a mobile web platform requiring no separate app installation. The seventh principle was exploring applicability and effectiveness through multidimensional data collection, triangulating survey results with system log data and applying a time-series design dividing the 10-week research period into existing learning and tool utilization periods.
Regarding technical implementation, the system was developed as a cross-platform mobile application operating smoothly on both mobile and web environments, considering accessibility across various user environments and long-term scalability and maintainability. The frontend adopted Google's Flutter framework, supporting iOS, Android, and Web simultaneously with a single codebase, while the backend was built on a serverless architecture enabling stable service without physical server management. Learning data was stored in Cloud Firestore, a NoSQL database supporting flexible storage methods optimized for storing hierarchically structured JSON data and synchronizing in real-time.
For vocabulary review, the HLR algorithm was applied to track learners' knowledge states. Log data generated whenever learners solved problems (correctness, response time, hint usage, etc.) was collected in real-time, and the system automatically analyzed the previous day's accumulated log data at 4 AM each morning via a scheduler to recalculate each learner's memory half-life for each vocabulary item. This process went beyond simple accuracy calculation to precisely estimate memory strength by incorporating various learning situation variables into the calculation formula. For grammar learning, the BKT model was applied, which assumes learners' knowledge state as a binary state of learned and not-learned and uses a Hidden Markov Model to infer this latent state. This study extended the classical BKT model to implement dynamic parameter updating logic reflecting learners' response speed, hint usage, and forgetting effects over time elapsed since learning.

The adaptive review engine synthesized these to design the overall customized learning list to be provided to learners, executed automatically by the server's scheduler during early morning hours (04:00) to avoid disrupting learner learning, proceeding through three stages of data aggregation, learning plan establishment, and difficulty scaffolding. Particularly, the system's most core feature was the adaptive difficulty scaffolding stage, where even for selected knowledge elements, problem types (difficulty) were differentiated according to learners' current mastery levels. In the vocabulary domain, recognition-focused matching exercises were presented at the initial stage when learners' memory probability was below 35%, fill-in-the-blank exercises to induce retrieval when memory probability rose to the 35%-65% range, and production tasks such as sentence making when sufficiently mastered above 65%. In the grammar domain, mechanical substitution exercises for learning morphological rules were assigned when grammar mastery probability was below 0.60, and response exercises requiring meaning-centered sentence generation were assigned when mastery exceeded 0.60, gradually increasing difficulty.
The learner interface was structured around three axes: learning assistant, dual-track review, and extended interaction to support learners' self-regulation and intrinsic motivation. The character "Suri," designed after Y University's symbol the eagle, functions as an educational agent throughout the system, appearing as a helper and friend from the beginning to end of learning, lowering the affective filter and creating psychological stability through friendly expressions and native language messages. The core review experience consisted of two complementary modules: Today's Review implementing external regulation (ER) and Adaptive Review supporting self-regulation (SR). Today's Review presents identical learning tasks to all class members according to schedules set by instructors, while Adaptive Review connects to personalized review lists generated by algorithms analyzing individual learners' learning histories.
The instructor dashboard was designed as a core tool systematically implementing the human-AI collaboration model. Through review plan management functions, instructors can directly set review content and start times to be applied to the entire class. The class-wide status board supports data-based instructional decision-making, allowing instructors to confirm not only average review completion and accuracy rates but also lists of common weak knowledge components where many learners made errors. Instructors can also select specific learners to view personalized review detailed status, error patterns, and changes in mastery by knowledge element estimated by HLR and BKT models. Writing monitoring and tutoring monitoring functions serve as devices ensuring AI's educational safety, allowing instructors to review writing feedback content or tutoring conversation records that learners had with generative AI.
To verify the tool's effectiveness, field application research was conducted over 10 weeks targeting 53 beginner-level (Level 1 and 2) learners enrolled in the 2025 fall semester regular course at Y University's Korean Language Institute and 3 instructors in charge. The research design distinguished an initial 5-week baseline establishment period from the subsequent 5-week tool utilization period, collecting and triangulating quantitative and qualitative data through midterm and final exam scores, system log data, surveys, and in-depth interviews.
Analysis of learning log data microscopically confirmed how AI algorithms actually operated. In the vocabulary domain, the stepped difficulty structure of matching (recognition), fill-in-the-blank (retrieval), and sentence making (production) supported learners' gradual progression from recognition to production, with the HLR algorithm accurately capturing repeated error items as weaknesses. In the grammar domain, the BKT algorithm distinguished and tracked mastery differences between substitution and response exercises. In adaptive review sessions, changes in learning patterns between consecutive sessions were observed, with significant improvement the next day on vocabulary items where errors occurred the previous day, confirming the possibility of distributed learning effects. However, consistency issues were discovered in evaluation functions using generative AI, with grammatically correct sentences being flagged as incorrect in sentence-making exercises, or diverse expressions conveying the same meaning not being accepted in response exercises, leading to cases where learners gave up after repeated attempts.
Furthermore, the Suri tutoring function demonstrated the possibility of AI-supported self-directed grammar learning. Although only one learner (L2-13) utilized this function, this learner conducted over 300 turns of in-depth conversation across 9 sessions, systematically exploring approximately 30 intermediate-to-advanced level grammar items. A metalinguistic inquiry pattern was observed where the learner consistently asked about tense combination possibilities, applicability to different verb types, and noun combinations for each grammar item, suggesting that AI tutors can support self-directed extended learning beyond the regular curriculum.
Analysis results showed stark polarization in learners' tool utilization patterns. Of 43 total registrants, only 10 were active users who consistently utilized the system throughout the research period, completing an average of 32.4 learning sessions per person, while the passive user group averaged only 3.0 sessions. In academic achievement analysis, both groups experienced grade declines due to increased difficulty in the latter half of the semester, but the active user group showed half the decline (-3.70 points) in listening compared to the passive user group (-6.53 points), suggesting the tool's protective effect on grades.
A notable finding of this study was the diagnostic validity of AI algorithms. Cross-analysis of the weak knowledge element lists generated by the system against actual final exam incorrect answers for 8 active users (Level 1) confirmed a total of 33 matching cases. Notably, the match rate was high in expressive domains such as writing (63.6%) and speaking (24.2%). An interesting point was that the nature of weaknesses captured by AI differed qualitatively according to learners' achievement levels. Upper-level learners showed fossilized errors such as particle misuse (confusion between '이/가' and '을/를') or overgeneralization of irregular conjugations (conjugating '묻다' as '물러'), while middle and lower-level learners primarily showed incomplete mastery of basic concepts such as failure to understand temporal sequence connective endings (-(으)ㄴ 후에) or conceptual confusion of basic particles.
However, when left to voluntary use, a distinct "rich-get-richer, poor-get-poorer" phenomenon emerged where only upper-level learners with already-established study habits actively utilized the tool while lower-level learners who actually needed additional support dropped out early. Instructor interviews also pointed out the phenomenon of only diligent upper-level students participating, with instructors acknowledging AI's potential while experiencing a dilemma between autonomy and compulsion, recognizing that "100% review ultimately requires the teacher to work one-on-one."
Survey data analysis based on the Technology Acceptance Model (TAM) identified key factors determining tool acceptance. The active user group consistently showed higher scores in prior behavioral intention (4.00 points) and self-regulated learning capability (3.81 points) compared to the passive user group (3.75 and 3.64 points respectively), confirming that learners with high intention for continued use and self-regulation capabilities before tool use were those who actually used it actively. Notably, while passive users showed higher perceived usefulness before use, a reversal occurred afterward with active users scoring higher, demonstrating that actual use experience strengthened usefulness perception.
Instructor interview analysis categorized experiences into four key themes: effectiveness and limitations of AI tools, workload and acceptance, learner characteristics and participation, and roles in the teaching-learning ecosystem. Instructors showed clear recognition that AI tools cannot completely replace instructor roles, with statements like "100% review ultimately requires the teacher to meet one-on-one" and "there are limits online for writing instruction which is very important for Level 1 students." Simultaneously, instructors experienced the tool's usefulness in specific situations, recognizing its value as an emergency replacement tool for filling learning gaps during instructor absence. The most prominent topic was the dilemma between autonomy and compulsion, with instructors consistently pointing out that autonomous tool effects heavily depend on learner diligence, concluding that "the app will only be used when the teacher takes initiative with some degree of compulsion."
These findings show that AI tools do not automatically produce educational effects. For AI's potential to be realized, active instructor participation and systematic external intervention for learners must be prerequisites. This study holds academic and practical contributions in being the first empirical research in the Korean language education field to apply HLR and BKT algorithms in a hybrid manner, presenting evidence of diagnostic validity where AI weakness diagnosis predicts actual language performance errors, and providing a concrete case of tool design organically linked to regular curriculum.
However, limitations exist including the small sample of active users (N=10), the short 5-week tool utilization period, and restriction to a single institution. Follow-up research requires expansion of research subjects, long-term longitudinal studies, and development of adaptive intervention strategies for lower-level learners. Furthermore, refinement research on generative AI evaluation functions is urgent, requiring systematic optimization of prompt engineering to resolve the sentence-making and response exercise consistency issues identified in this study's log analysis.
In conclusion, this study demonstrated the possibility that AI utilization research in Korean language education can move beyond the exploration stage to the empirical stage of actually developing tools and applying them in the field. However, the "rich-get-richer" effect, generative AI evaluation consistency issues, early dropout due to technical friction, and the autonomy-compulsion dilemma discovered in this study serve as important guidelines for introducing AI tools in Korean language education settings. Ultimately, this study demonstrated the potential for collaboration between AI and instructors within regular Korean language curriculum, while clearly showing that for such collaboration to work effectively, learner tool use must precede, which in turn requires active instructor participation and external intervention for learners.

국문 초록 (Abstract)

본 연구는 대학 부설 한국어 교육기관의 초급 학습자를 대상으로 AI 기반 맞춤형 복습 도구를 설계·개발하고 이를 정규 교육과정에 적용하여 교육적 효과와 현장 적용 가능성을 검증하고자 하였다. 연구의 출발점은 한국어 학습 수요의 급증에도 불구하고 다양한 배경의 학습자가 혼재된 초급 교실에서 개인화된 복습 기회를 제공받기 어려운 구조적 한계에 대한 문제의식이었다.
본 연구는 자기조절학습(SRL), 자기결정성이론(SDT), 간격효과, 지식 추적 등 검증된 교육 이론을 토대로 7가지 핵심 설계 원칙을 도출하였다. 특히 언어 지식의 인지적 특성에 주목하여 어휘에는 망각을 모델링하는 반감기 회귀(HLR) 알고리즘을, 문법에는 지식 습득 여부를 추적하는 베이지안 지식 추적(BKT) 알고리즘을 차별적으로 적용하는 하이브리드 접근을 시도하였다. 또한 AI가 교수자를 대체하는 것이 아니라 데이터로 지원하는 협업 파트너가 되도록 교수자 대시보드를 핵심 기능으로 설계하였다.
개발된 도구는 Y대학교 한국어 교육기관의 초급(1급, 2급) 학습자 53명과 담당 교수자 3명을 대상으로 10주간 적용되었다. 분석 결과, 학습자들의 도구 활용 양상은 극명한 양극화를 보였다. 학습자 53명 중 43명이 가입하였고 그중 적극적 사용자는 10명에 불과하였다. 적극적 사용자는 1인당 주 평균 32.4회의 학습 세션을 완료한 반면 소극적 사용자 집단은 1인당 주 평균 3.0회로 10배를 넘는 차이를 보였다. 중간고사 대비 기말고사의 성적 변화(기말-중간)를 분석한 결과, 전체적인 점수 하락 추세 속에서도 적극적 사용자 집단은 듣기 영역에서 소극적 사용자 집단(-6.53점)에 비해 절반 수준의 하락폭(-3.70점)을 보여 도구의 성적 하락 방어 효과가 시사되었다.
본 연구의 주목할 만한 발견은 AI 알고리즘의 진단 타당성이다. 적극적 사용자 8명을 대상으로 시스템이 산출한 취약 지식 요소 목록과 실제 기말고사 오답을 교차 분석한 결과, 총 33건의 일치 사례가 확인되었다. 특히 학습자의 성취도 수준에 따라 AI가 포착한 취약점의 성격이 질적으로 달랐다. 상위권 학습자에게서는 조사 오용이나 불규칙 활용의 과잉 일반화와 같은 화석화된 오류가 관찰된 반면, 중·하위권 학습자에게서는 기초 개념 미숙달이 주된 연계 요인으로 나타났다.
또한 AI 튜터 수리와의 상호작용에서는 학습자가 문법 항목에 대해 시제 결합 가능성, 동사 유형별 적용 등을 체계적으로 질문하는 메타언어적 탐구 패턴이 관찰되어, AI 튜터가 정규 교육과정을 넘어선 자기주도적 확장 학습을 지원할 수 있는 가능성이 확인되었다.
그러나 자발적 사용에 맡겼을 때 이미 학습 습관이 형성된 상위권 학습자만 도구를 적극 활용하고 하위권 학습자는 초기에 이탈하는 부익부 빈익빈 현상이 뚜렷하게 나타났다. 이는 AI 도구가 저절로 교육적 효과를 발휘하는 것이 아니며 교수자의 적극적인 참여와 학습자에 대한 체계적인 외부 개입이 전제되어야 함을 시사한다.
본 연구는 한국어교육 분야에서 HLR과 BKT 알고리즘을 하이브리드로 적용한 실증 연구라는 점, AI의 취약점 진단이 실제 언어 수행 상의 오류를 예측한다는 진단 타당성의 증거를 제시한 점, 그리고 정규 교육과정과 유기적으로 연계된 도구 설계의 구체적 사례를 제시한 점에서 학문적·실천적 기여를 지닌다.

목차 (Table of Contents)

• 그림 차례 ⅳ
• 표 차례 ⅵ
• 부록 차례 ⅶ
• 국문 요약 ⅷ
• 제1장 서론 1
• 제2장 이론적 배경 6
• 2.1. 학습자의 자기주도성과 지속성 6
• 2.1.1. 자기조절학습 6
• 2.1.2. 자기결정성 이론 9
• 2.2. 인지적 효율성을 위한 학습 원리 11
• 2.2.1. 모바일 러닝 12
• 2.2.2. 마이크로러닝 13
• 2.2.3. 분산학습과 간격효과 14
• 2.3. 디지털 학습 환경과 기술 수용 17
• 제3장 적용 기술 및 선행 연구 20
• 3.1. 지식 추적 알고리즘 20
• 3.1.1. 베이지안 지식 추적(BKT) 20
• 3.1.2. 딥러닝 기반 지식 추적의 발전 22
• 3.1.3. 반감기 회귀(HLR) 알고리즘 23
• 3.2. 생성형 AI의 교육적 활용 25
• 3.3. 지속 가능한 생태계를 위한 인간-AI 협업 29
• 3.4. 선행연구 종합 32
• 제4장 AI 복습도구의 설계 및 개발 34
• 4.1. AI 복습도구의 설계 원칙 35
• 4.2. 시스템의 개념적 틀 36
• 4.3. 교수·학습 맥락 및 데이터 설계 38
• 4.4. 시스템 구현 43
• 4.4.1. 시스템 아키텍처 및 기술 스택 43
• 4.4.2. HLR 기반 어휘 학습 상태 추적 44
• 4.4.3. BKT 기반 문법 학습 상태 추적 47
• 4.4.4. 적응형 복습 생성 로직 51
• 4.5. 사용자 인터페이스 설계 53
• 4.3.1. 학습자 인터페이스 53
• 4.3.2. 교수자 인터페이스 65
• 제5장 현장 적용 연구 방법 70
• 5.1. 연구 설계 및 참여자 70
• 5.2. 연구 절차 71
• 5.3. 측정 도구 및 데이터 수집 72
• 5.4. 데이터 분석 73
• 제6장 현장 적용 결과 분석 75
• 6.1. 학습 로그를 통해 본 학습자의 복습 경험 76
• 6.1.1. 어휘 연습 76
• 6.1.2. 문법 연습 83
• 6.1.3. 대화 연습 90
• 6.1.4. 자유 작문 연습 91
• 6.1.5. 수리 튜터링을 통한 개별화된 질의응답과 상호작용 95
• 6.1.6. 매일 복습 99
• 6.1.7. 맞춤형 복습 101
• 6.2. 학습 도구의 사용 양상과 기술 수용 과정 107
• 6.2.1. 자발적 수용에 따른 사용자 집단 유형화 107
• 6.2.2. 시계열적 사용 추이 및 이탈 시점 분석 108
• 6.2.3. 기술적 마찰과 초기 수용 장벽 109
• 6.3. 앱 사용 경험과 학업 성취도 간의 관계 110
• 6.4. AI 알고리즘의 진단 및 예측 타당성 검증 113
• 6.5. 현장 수용성 및 교육적 시사점 분석 117
• 6.5.1. 학습자의 기술 수용과 사용 경험 117
• 6.5.2. 교수자의 도구 수용 양상과 저해 요인 분석 121
• 제7장 결론 및 제언 125
• 7.1. 연구 요약 125
• 7.1.1. 교육적 설계 원칙과 시스템 구현 125
• 7.1.2. 학습자 경험과 학업 성취 126
• 7.1.3. 교수자 지원과 인간과 AI의 협업 129
• 7.1.4. 지속 가능한 AI 활용 모델과 제언 130
• 7.2. 연구의 의의 132
• 7.3. 연구의 한계 및 제언 135
• 참고 문헌 141
• 부록 150
• 영문 요약 205