딥러닝 기반 단어 임베딩을 적용한 사진 자막 영작문 채점 시스템
Automatic Scoring System for Picture-based English Caption Writing Test Adopting Deep Learning Based Word-Embedding
- 주제(키워드) Key Words: computer assisted language learning , deep learning , English writing , word embedding , criterion-referenced test , scoring , 주제어(Key Words): 컴퓨터 언어보조학습(computer assisted language learning) , 딥러닝(deep learning) , 영작문(English writing) , 단어 임베딩 (word embedding) , 준거참조검사(criterion-referenced test) , 채점(scoring)
- 주제(기타) 언어학
- 설명문(URI) https://www.kci.go.kr/kciportal/ci/sereArticleSearch/ciSereArtiView.kci?sereArticleSearchBean.artiId=ART002739767
- 등재 KCI등재
- 발행기관 대한언어학회
- 발행년도 2021
- URI http://www.dcollection.net/handler/ewha/000000182854
- 본문언어 한국어
- Published As http://dx.doi.org/10.24303/lakdoi.2021.29.2.1
초록/요약
Kim, Dongsung. (2021). Automatic scoring system for picture-based English caption writing test adopting deep learning based word-embedding. The Linguistic Association of Korea Journal, 29(2), 1-20. Since human grading of English writing requires substantial resources, many researchers in the area of Computer-Assisted Language Learning (CALL) have been focusing on automatic scoring systems based on natural language processing systems, machine learning, and other automatic processing mechanisms. English Testing Services (ETS) announced several automatic scoring systems for English writing. In this paper, we suggest using a deep learning based automatic scoring system for an English caption writing test. Our method involves using a sentence similarity measurement, which compares different levels of answer sentences with user writing input. We chose different word embedding types (Word2Vec, Word Mover‘s Distance (WMD), Bidirectional Encoder Representations from Transformers (BERT)) and Abstract Meaning Representation (AMR), a linguistic model for comparing semantic differences between two sentences based on semantic representation. Scoring systems should not only satisfy the requirements of complicated scoring rubrics but also meet the conditions of a language proficiency test. Our results show that BERT outperforms three competitive models in predicting accurate scoring levels and also shows the characteristics of the criterion reference which could theoretically express the standards of a language proficiency test.
more