소셜미디어 텍스트의 슬랭 처리 기반 계층적 감정 분류와 텍스트 감정 맞춤형 이모지 생성 기법

초록

This paper proposes a user-tailored emoji generation method grounded in fine-grained emotion analysis of social-media text. We reorganize GoEmotions into a two-level hierarchy (7 superclasses, 27 subclasses) and design a BERT-based coarse-to-fine classifier with a shared encoder and grouped heads. An OED-based slang dictionary with normalization improves robustness to nonstandard expressions. We use the classifier outputs to condition DALL·E 3 via three prompt strategies—S1 (text), S2 (text, emotion), and S3 (text, emotion, Description of emotion). In our experiments, the hierarchical model with slang normalization produced the best F1, and S3 achieved the highest CLIP-Sentiment and CLIP-Text, with no IS degradation, while maintaining image quality and diversity.

키워드