L2 발화 유창성 지표와 학습자 변인 간의 관계에 대한 머신러닝 연계 분석
A Machine Learning–Integrated Analysis of Relationships Between L2 Speech Fluency Indices and Learner Variables
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
This study examines utterance fluency in advanced L2 Korean learners from Chinese and Vietnamese L1 backgrounds, relative to native speakers, using a multidimensional framework spanning speed, pauses, repairs, and prosody. Twenty-one native speakers and sixty-two TOPIK Level 5–6 learners completed a story-retelling task. Speech was analyzed for articulation rate (AR), intra-clausal pause ratio (ICPR), mean length of run (MLR), total pause ratio (TPR), self-repair/repetition rate (SRR), and pitch dynamism quotient (PDQ). Independent-samples tests, MANCOVA with FDR correction, and machine-learning classification with nested cross-validation (logistic regression, random forest) were employed. Group comparisons showed significant differences in ICPR, MLR, SRR, and TPR, whereas AR and PDQ did not reliably differ between natives and learners. Within the learner group, only AR consistently reflected both L1 background and proficiency. In classification, both models exceeded a majority-class baseline, achieving roughly 90–93% out-of-sample accuracy. Permutation-importance analyses indicated that ICPR accounted for most predictive value (M = 66.7%, SD = 27.8), followed by AR (M = 19.7%, SD = 18.6) and MLR (M = 10.8%, SD = 15.4); TPR, SRR, and PDQ together contributed minimally. A shallow surrogate tree visualizing random-forest decision tendencies split first on ICPR, with AR and MLR providing secondary cues. Findings suggest that, at advanced proficiency, pause location and speech continuity are more discriminative than overall speed, motivating increased weight on ICPR in assessment and instruction. Results also align with possible cross-linguistic rhythm/tempo transfer among tonal-language learners and support integrative fluency assessment that prioritizes location-based pause measures complemented by speed-related indices.
Abstract
This study examines utterance fluency in advanced L2 Korean learners from Chinese and Vietnamese L1 backgrounds, relative to native speakers, using a multidimensional framework spanning speed, pauses, repairs, and prosody. Twenty-one native speakers and sixty-two TOPIK Level 5–6 learners completed a story-retelling task. Speech was analyzed for articulation rate (AR), intra-clausal pause ratio (ICPR), mean length of run (MLR), total pause ratio (TPR), self-repair/repetition rate (SRR), and pitch dynamism quotient (PDQ). Independent-samples tests, MANCOVA with FDR correction, and machine-learning classification with nested cross-validation (logistic regression, random forest) were employed. Group comparisons showed significant differences in ICPR, MLR, SRR, and TPR, whereas AR and PDQ did not reliably differ between natives and learners. Within the learner group, only AR consistently reflected both L1 background and proficiency. In classification, both models exceeded a majority-class baseline, achieving roughly 90–93% out-of-sample accuracy. Permutation-importance analyses indicated that ICPR accounted for most predictive value (M = 66.7%, SD = 27.8), followed by AR (M = 19.7%, SD = 18.6) and MLR (M = 10.8%, SD = 15.4); TPR, SRR, and PDQ together contributed minimally. A shallow surrogate tree visualizing random-forest decision tendencies split first on ICPR, with AR and MLR providing secondary cues. Findings suggest that, at advanced proficiency, pause location and speech continuity are more discriminative than overall speed, motivating increased weight on ICPR in assessment and instruction. Results also align with possible cross-linguistic rhythm/tempo transfer among tonal-language learners and support integrative fluency assessment that prioritizes location-based pause measures complemented by speed-related indices.