DeepASD Framework: A Deep Learning-Assisted Automatic Sarcasm Detection in Facial Emotions

Abstract

The vibrant human-machine research provides space for assessing sentiments in facial emotions. Capturing apt sarcasm-related emotions, especially in online meetings or stress interviews, is a challenging aspect. The purpose of this research is to apply deep learning algorithms to effectively assess the sarcasm in human facial emotions in an automatic fashion using the proposed Deep Learning-Assisted Automatic Sarcasm Detection (DeepASD) framework. Our framework trains facial sarcasm-related emotions from internet sources and applies deep learning algorithms to perform visual sarcasm detections. The proposed framework processes algorithms on edge-enabled compute nodes, including GPU-based machines. We evaluated the DeepASD framework using various deep learning algorithms such as EfficientNet, XceptionNet, InceptionNet, ResNet, DenseNet, ConvNext, MobileNet, and their variants; and, we observed that Mobilenetv3 achieved a better learning accuracy of 96.44 percent and energy consumption of 7959 Joules using minimal trainable/non-trainable parameters while detecting sarcasm in facial emotions. Our work will be beneficial for online interviewers, business enthusiasts, or future robotic machine developers for accomplishing accurate decisions considering sarcasm in facial emotions.