Abstract
Approximately 20% of the world’s population suffers from mental health disorders. Despite this, resources for mental health around the world remain scarce, inequitable, and inefficient. With the rapid development of information technologies, conversational agents have been proposed as a solution to increase the accessibility of mental health services and reduce the workload of psychotherapists. In recent years, a growing body of research has focused on what social skills or psychotherapy strategies these conversational agents should consider when providing mental health support. It is also shown that using emotional support strategies can better comfort patients and when used at the right time patients’ problems can be well identified. Although the importance of emotional support strategies has been demonstrated, how to use these strategies appropriately in the dialogue remains a major challenge. To address this issue, our goal is to employ deep learning technologies to help determine which strategy to use in an emotional support conversation. In this paper, we construct a task aiming at predicting emotional support strategies in a dialogue. In our model, in addition to using BERT and Bi-LSTM for integrating the dialogue history, we further include three supporter-related features from the dialogue. Through extensive experiments, we demonstrate the effectiveness of our method and show that the supporter-related features improve the Macro F1 scores by almost 3% which are critical for the prediction.