EMOCOV: Machine learning for emotion detection, analysis and visualization using COVID-19 tweets.

Area of Science:

• Computational Social Science
• Natural Language Processing
• Mental Health Research

Background:

• The Covid-19 pandemic significantly impacted global mental health, necessitating analysis of public emotional responses.
• Understanding fine-grained emotions like anger and fear in public discourse is crucial for mitigating socio-economic damage.

Purpose of the Study:

• To develop and evaluate a neural network model for automatic detection of fine-grained emotions in Covid-19 related tweets.
• To create a novel dataset of manually labeled tweets capturing emotional responses during the pandemic.
• To develop a custom Question-Answering (Q&A) RoBERTa model for extracting emotion-driving phrases from tweets.

Main Methods:

• A manually labeled dataset of Covid-19 tweets and general tweets was created.
• A neural network classification model was trained on this dataset for emotion detection.
• A custom RoBERTa Q&A model was developed to identify specific phrases linked to emotions.

Main Results:

• The classification model achieved an accuracy of 0.8951 and a Jaccard score of 0.6475.
• The custom RoBERTa Q&A model demonstrated superior performance with a Jaccard score of 0.7865.
• The study provides a historical analysis of emotions expressed in US Covid-19 tweets, including state-level insights.

Conclusions:

• The developed models effectively detect emotions and identify their sources in Covid-19 tweets.
• This work contributes a unique dataset and methodology for analyzing public mental health during crises.
• The findings offer valuable insights into the emotional landscape of the pandemic and its societal impact.