인문학
사회과학
자연과학
공학
의약학
농수해양학
예술체육학
복합학
DBpia AI
AI 에이전트
AI 아이디어
AI 리더
AI 뷰어
Citeasy
크롬 서지관리 다운로드
지원사업
학술연구/단체지원/교육 등 연구자 활동을 지속하도록 DBpia가 지원하고 있어요.
커뮤니티
연구자들이 자신의 연구와 전문성을 널리 알리고, 새로운 협력의 기회를 만들 수 있는 네트워킹 공간이에요.
초록· 키워드
In this study, we evaluated the humor comprehension abilities of Large Multimodal Models (LMMs) by using images and captions. We examined language and model differences based on the accuracy of humor selection through multiple-choice questions. As far as English is con- cerned, publicly available datasets were utilized. But, in case of Korean, we manually curated image-text pairs on ‘Jemok Hakwon’, which features cleverly captioned images. We inves- tigated whether LMMs could select humorous captions relevant to the image or reject irrele- vant ones. Language comparison between English and Korean revealed poorer performance for Korean, which is attributed to cultural elements based on error analysis. In the mod- el-by-model comparison between GPT-4V and Claude 3, GPT-4V exhibited superiority. But both models performed in the low 50-60% range. This relatively low performance indicates that humor comprehension, demanding high cognitive abilities, is still challenging for artifi- cial intelligence. This study is significant in that it constructs and publicly releases a Korean humor dataset based on Image-Text, enabling quantitative evaluation of state-of-the-art LMMs’ humor comprehension across languages and models. Through this study, we aim to stimulate further research measuring the upper limit of artificial intelligence using abstract information.
상세정보 수정요청해당 페이지 내 제목·저자·목차·페이지정보가 잘못된 경우 알려주세요!
목차
등록된 정보가 없습니다.
참고문헌
참고문헌 신청최근 본 자료
