EmpatheticExchanges: Toward Understanding the Cues for Empathy in Dyadic Conversations

This article presents a new research task and dataset to understand empathetic responses in conversations between two people. We formulate the empathy prediction problem as a way to model how people with different empathetic abilities react to a dialog by another person. Our novel approach focuses on dyadic conversations within positive and negative emotional contexts. To tackle the proposed problem, we introduce a new dataset, EmpatheticExchanges, presenting the largest sample size of manually annotated exchanges. Besides, EmpatheticExchanges includes empathy features validated by state-of-the-art research in affective computing, such as VAD vector values. Finally, we provide a baseline pattern-based model with a Closeness Evaluation measure score of 0.597 and a weighted F1 score of 0.452 using a 3-point scale. This model is explainable and performs similarly to popular techniques such as deep learning. Our results validate the quality of our dataset and the use of a pattern-based approach for empathy classification.