한국어 자연어 처리 모델 간 도메인별 감성 분류 성능 비교 분석 = Comparative Analysis of Emotional Classification Performance by Domain between Korean Natural Language Processing Models

국문 초록 (Abstract)

최근 AI 기술의 발달로 IT 기업에서는 AI를 활용한 문제 해결 및 서비스 연구개발 움
직임이 급증하였다. 반면 서비스 가능 수준의 성능을 지닌 AI 모델을 구축하기 위해서는
문제 상황의 도메인을 고려하여 데이터 수집, 가공, 검수되어야 하며, 전문가의 반복적인
모델 최적화, 실험과 같은 시간적, 인력적으로 높은 비용이 발생할 수 있는 과정을 거쳐야
한다. 본 논문에서는 투자비용 대비 높은 성능 확보에 유리한 공개된 한국어 사전 학습 언
어 모델 8종을 활용하여 감성 분석 태스크에서의 분류 성능을 비교 분석하고, 다양한 도메
인에 대하여 범용적으로 좋은 성능을 보여주는 언어 모델을 확인하고자 한다. 이를 위해
채택한 8종의 사전 학습 언어 모델을 영화 도메인 데이터로 전이 학습한다. 그리고 수집한
6종의 도메인별 한국어 리뷰 데이터를 각 모델에 교차 검증하고, 통계적 검정을 통해 모델
간 성능을 비교한다. 결과적으로 학습된 영화 도메인 외 다른 도메인에서도 좋은 성능을
보여주는 범용성이 높은 모델은 LMkor-BERT임이 확인되었다. 본 논문을 통해 한국어
사전 학습 언어 모델 채택 시 비용적인 부분에서 보다 효율적인 베이스라인 모델을 선택
할 수 있는 기초자료로 활용될 수 있다.

목차 (Table of Contents)

• 목 차
• 국문요약
• 1. 서 론 ·············································································································2
• 1.1. 연구 배경 ·······························································································2
• 1.2. 연구 목적 ·······························································································3
• 1.3. 실험 과정 ·······························································································3
• 2. 관련 연구 ··················································································4
• 3. 감성 분석 모델 ········································································5
• 3.1. BERT ·····························································································5
• 3.2. ALBERT ························································································9
• 3.3. ELECTRA ·····················································································9
• 3.4. FUNNEL ·······················································································10
• 4. 모델 구현 ··················································································10
• 4.1. 데이터셋 설명 ··············································································10
• 4.2. 모델 전이 학습 정보 ··································································11
• 5. 모델 성능 결과 ····················································································· 11
• 5.1. 정규성 검정 ························································································· 13
• 5.2. Levene 등분산 검정 ········································································ 14
• 5.3. Anova 일원 분산 분석 ··································································· 14
• 5.4. 쌍체 t-검정 ························································································ 14
• 6. 결론 ············································································································· 15
• 참고문헌 ·········································································································· 16