Feature Engineering Trends in Text-Based Affective Computing: Rules to Advance Deep Learning Models

Abstract

Understanding emotions in textual data, particularly within dynamic social media platforms such as YouTube, Facebook, and Twitter, presents significant challenges. This paper aims to provide a comprehensive review of emotion detection techniques in affective computing, highlighting key advancements, challenges, and ethical concerns. The key contributions of this review include an examination of foundational theories of NLP-based emotion recognition, an analysis of the role of affect lexicons in emotional classification, and a review of commonly used datasets for training emotion detection models. Additionally, it explores various feature extraction techniques, including lexicon-based approaches such as SentiWordNet and NRC Emotion Lexicon, statistical and syntactic features like n-grams and POS tags, and semantic embeddings from deep learning models such as Word2Vec, GloVe, BERT, RoBERTa, and GPT. Findings show that while deep learning and transformer models improve contextual understanding, they also introduce challenges such as high computational costs, data imbalance, and domain adaptability issues. Bias in training data poses ethical risks, potentially reinforcing stereotypes and enabling manipulative applications like targeted advertising and misinformation. Key research gaps include the need for improved feature representations, bias mitigation, enhanced model accuracy and fairness. Traditional models struggle with real-world complexities, while transformer-based models face challenges related to scalability, dataset limitations, and interpretability. Addressing these challenges will enhance affective computing accuracy, fairness, and applicability across industries such as healthcare, education, and human-computer interaction.