The task of detecting emotions from text has become increasingly popular in recent years, mostly due to significant recent advances in Natural Language Processing (NLP). Emotion recognition or classification from text is a challenging task, as emotions are hard to describe or pinpoint, and different emotions such as joy and anxiety can co-occur, especially for longer text records. One of the issues with text understanding in general, but emotion recognition in particular, is the small size of available data that has been annotated with emotions by experts. Our literature survey shows that researchers have tried to alleviate this issue by artificially generating text in order to enlarge or augment the data. This can be achieved in various different ways, for example replacing a word with a synonym or paraphrasing the existing text using chatGPT, etc. In this study, we continue our previous work in emotion detection with small text datasets. Our goal is to investigate different data augmentation techniques on our data and especially their impact on the model performance given the task of emotion recognition. We utilize state-of-the-art machine learning models to detect emotions based on textual comments and perform extensive experiments to compare the effects of the various augmentation approaches.
- COVID-19 Survey: collected via survey in UK in lockdown April 2020. 2,500 participants wrote text responses and rated their emotions e.g. worry, anxiety, relaxation, etc. https://github.com/ben-aaron188/covid19worry
- WASSA: part of the WASSA 2021 Shared Task. Contains essays written after reading news articles related to harm to an individual, nature, etc. https://competitions.codalab.org/competitions/28713
- EmoEvent: dataset composed of tweets from the Twitter platform whcih exhibit different emotions in relation to various events. https://github.com/fmplaza/EmoEvent