딥러닝 언어모델의한국어 학습자 말뭉치 원어민성 판단 결과 분석 연구
A Study on the Judgment of Nativelikeness of Korean Learner Corpus by Deep Learning Language Model.
0
조회
0
다운로드
7
인용
0
참고문헌
키워드
초록
The present study aims to analyze how deep learning judges nativelikeness in the corpus of native Korean speakers and Korean learners. To this end, a deep learning model that classifies sentences of native Korean speakers and Korean learners was built, and the criteria for determining nativelikeness between deep learning and humans were compared and analyzed in terms of error analysis. As a result of the analysis, the accuracy of the deep learning model built in this study was found to be 91%, which means that 91 sentences out of 100 sentences were accurately classified whether they were written by the native speaker or by the learner. In addition, since the error annotation result of the learner corpus is a projected result of human judgment of nativelikeness, the similarities and differences of the criteria for determining nativelikeness were described in detail by comparing it with the test data verification result of deep learning. The results of this study will be an important basis for clarifying what the nativelikeness of native Korean speakers is and for objectively judging the nativelikeness of the language produced by Korean learners.
Abstract
The present study aims to analyze how deep learning judges nativelikeness in the corpus of native Korean speakers and Korean learners. To this end, a deep learning model that classifies sentences of native Korean speakers and Korean learners was built, and the criteria for determining nativelikeness between deep learning and humans were compared and analyzed in terms of error analysis. As a result of the analysis, the accuracy of the deep learning model built in this study was found to be 91%, which means that 91 sentences out of 100 sentences were accurately classified whether they were written by the native speaker or by the learner. In addition, since the error annotation result of the learner corpus is a projected result of human judgment of nativelikeness, the similarities and differences of the criteria for determining nativelikeness were described in detail by comparing it with the test data verification result of deep learning. The results of this study will be an important basis for clarifying what the nativelikeness of native Korean speakers is and for objectively judging the nativelikeness of the language produced by Korean learners.
참고문헌 (6)
- 고언숙, "정도부사 ‘너무’의 의미와 의미운율: 음성 말뭉치 분석을 바탕으로", 음성음운형태론연구, 21(2), pp.223~243 (2015), 한국음운론학회[DOI]
- 동뢰, "확장 의미 단위(EUM) 기반 한국어 어휘 교육 방안 탐구", 외국어로서의 한국어교육, 56, pp.105~132 (2020), 언어연구교육원 한국어학당[DOI]
- 박권식, "한국인 영어학습자의 영어 문장은 얼마나 원어민스러운가: 딥러닝 기반 분석", 언어사실과 관점, 48, pp.195~227 (2019), 언어정보연구원[DOI]
- 2019년 한국어 학습자 말뭉치 연구 및 구축 사업, "한송화", 국립국어원, ~ (2019)
- Devlin, J., "Bert: Pre-training of deep bidirectional transformers for language understanding", arXiv preprint arXiv:1810.04805, ~ (2018)
- 박권식, "Not Yet as Native as Native Speakers: Comparing Deep Learning Predictions and Human Judgments", 영어학연구, 26(1), pp.199~228 (2020), 한국영어학학회[DOI]