Enhancing Emotion Detection through ChatGPT-Augmented Text Transformation in Social Media Text
Sanghyub John Lee, Hyunseo Tony Lee, Ki-Seong Lee · 2024
Social networking services (SNS) provide a rich source of user-generated emotion-expressed text. However, deciphering the emotion from these texts, often marked by vernacular expressions and abbreviations, poses significant challenges. This study introduces a novel approach to enhance emotion detection by converting SNS informal texts into everyday language using ChatGPT, resulting in a generative large language model (LLM). The study uses ten publicly available emotion datasets and a unique tweet dataset, augmented through ChatGPT. Three models were trained for comparison: one using original texts (n=408,359), another with ChatGPT-augmented texts (n=408,359), and the last with a combination of both (n=816,718). The four transformer models, RoBERTa, BERT, DistilBERT, and XLM-RoBERTa, trained with the combined dataset outperformed those trained solely on original texts, indicating that converting SNS vernacular text into everyday language improves emotion detection. The study provides significant insights for enhancing emotion analysis in human-robot interaction and other fields reliant on accurate emotion detection, demonstrating the potential of LLMs in natural language processing (NLP) data augmentation.