Context-Dependent Deep Learning for Affective Computing

Deep-learning models have been widely employed for recognizing emotions from various modalities. Yet these models face a number of challenges such as generalizing to different test conditions and predicting fine-grained emotions to name a few. One possible way to tackle these challenges is to provide these deep-learning models with additional context which can be in the form of domain knowledge from external knowledge sources or inherent task properties in the form of task-specific auxiliary losses. We hypothesize that incorporating context can better guide deep-learning models to look at the right features. In this extended abstract, we specifically focus on the problem of fine-grained emotion recognition using text-based data. We explore how to augment state-of-the-art NLP models with context to improve their performance in detecting fine-grained classes. We also discuss the implications of our research and future directions.