<![CDATA[Reaction images are a form of visual communication widely used in social media comments, including TikTok. They can convey direct responses or indirect communicative intentions, such as sarcasm, irony, evaluation, symbolism, and implicit humor. This study aimed to develop and validate a classification of communicative intentions into Literal and Figurative Intentions and implement a Convolutional Neural Network (CNN) using EfficientNetB0 transfer learning. A total of 339 images were annotated by three annotators using an expert-judged guideline. After revising the guideline and relabeling 124 disagreements, the average descriptive Cohen’s Kappa increased from 0.389 to 0.886. Following duplicate removal, 334 images remained, comprising 245 Literal and 89 Figurative Intentions. The data were stratified into training, validation, and test sets. The baseline achieved 76.47% accuracy, 67.31% macro F1-score, 75.45% weighted F1-score, and 0.6933 AUC. Figurative Intentions remained more difficult to classify, indicating limitations of unimodal visual approaches for context-dependent communication.]]>
Copyrights © 2026
