ACM/SIGAPP SAC 2025 AIED accepted paper (Oral) Paper Arxiv
- Accurate segmentation of Korean morphemes
- Challenging due to agglutinative nature (frequent morphological changes)
- Errors propagate and negatively affect higher-level analyses
- Utilize a state-of-the-art Korean morpheme analyzer
- Minimize errors in morpheme analysis
- Morpheme analyzer: Bareun
- Morpheme analyzer used for vocabulary grading: UTagger
- Extract diverse linguistic features from morpheme level to sentence, paragraph level features
- Over 294 numerical features, categorized as
- Basic features: morpheme counts, density, lengths
- Lexical diversity:
- Type-Token Ratios (TTR, RTTR, CTTR)
- MSTTR, MTLD, HD-D, VocD
- Cohesion features: semantic similarity, topic consistency, etc.
- Produce explainable, rubric-based writing scores
- Predict 10 rubric scores per essay using attention-based deep learning model
| N | Type | Rubric |
|---|---|---|
| 1 | 표현 (Expression) | 문법 (Grammar) |
| 2 | 어휘 (Vocabulary) | |
| 3 | 문장 표현 (Sentence Expression) | |
| 4 | 구조 (Organization) | 문단 내 구조 (In-paragraph Structure) |
| 5 | 문단 간 구조 (Inter-paragraph Structure) | |
| 6 | 구조적 일관성 (Structural Consistency) | |
| 7 | 길이 (Length) | |
| 8 | 내용 (Content) | 주제 명확성 (Topic Clarity) |
| 9 | 독창성 (Originality) | |
| 10 | 서사 (Narrative) |
- Combines
- Sentence-level representations (contextual meaning via pre-trained LM + BiGRU)
- Essay-level features (lexical and cohesion metrics)
- Explainability through attention
- Identifies which essay-level features most influence final scores
- Provides transparency and reliability to users