Data augmentation using instruction-tuned models improves emotion analysis in tweets

With ever-growing internet users, Twitter has evolved as one of the most popular social media platforms for voicing opinions. Emotion analysis in tweets is helpful in many applications. ChatGPT, the recent AI buzz, received significant attention for its extraordinary abilities. The effectiveness of ChatGPT has not yet been evaluated for emotion analysis in tweets. This paper investigates how effective ChatGPT is for tweet emotion analysis. Experimental results show that directly applying ChatGPT in zero or few-shot settings does not yield good results. ChatGPT is more effective for data augmentation, and fine-tuning smaller pre-trained models on the augmented dataset achieves new state-of-the-art results on the TweetEval emotion analysis dataset. It is the first work to show that ChatGPT augmented tweets can help smaller pre-trained models achieve the best results for tweet emotion analysis.