Emotion Detection from Real-Life Situations Based on Journal Entries Using Machine Learning and Deep Learning Techniques

Emotion regulation is crucial for psychological well-being. Negative emotions such as anger, fear, and sadness have been shown to create unhealthy patterns of physiological functioning and reduce human resilience and quality of life. Positive emotions (e.g., happiness), however, improve physical and mental health. Therefore, our goal is to: (1) build emotion detection models trained on a dataset that captures people’s real-life situations and experiences using machine learning (ML) and deep learning techniques, and then (2) design and develop an emotion-adaptive and persuasive mobile health (mHealth) application that delivers interventions for coping with challenging life events, promote positive emotions, and motivate behaviour change for improved mental health and well-being. Current work is focused on the first goal in which we trained and evaluated twenty-one classical ML, ensemble learning, and deep learning models to determine their capability in detecting five popular emotional states: anger, fear, disgust, happiness, and sadness from journal entries. The overall best performing model is MCBiLSTM (F1-score = 81.1%), a multichannel fusion model comprising three Convolutional Neural Network (CNN) channels and a Bidirectional Long Short-Term Memory (BiLSTM) recurrent neural network. Future work will integrate the model into our proposed mHealth application (second goal) to tailor therapeutic interventions based on individuals’ current emotional state in real-time.