Producing Considerate Responses: Progressive Staged Training for Emotional Support Conversation

Emotional support conversation (ESC) aims to alleviate the negative emotions of help-seekers by providing psychological assistance. Existing approaches typically overlook the abundant annotations contained in the ESC dataset, such as the situation descriptions and feedback scores of seekers, which limits their performance. In an effort to utilize the annotation information to enhance the emotional support ability of the backbone, we propose a three-stage training method called BlenderBot-ThTra for ESC systems. The proposed BlenderBot-ThTra involves the following three training processes: fine-tuning with supplemental feedback utterance, fine-tuning with auxiliary situation restoration, and calibration with the helpfulness estimation. The first stage aims to intensify the backbone's perception of conversational context, the second stage propels the backbone into excavating the causes of the emotional distress faced by the seeker. In the third stage, we leverage a Bayesian method based on the seeker's feedback scores to train a helpfulness evaluation model, then exploit a contrastive learning method to calibrate the ESC backbone. We conduct experiments on the standard multiturn ESC dataset, and the results demonstrate that BlenderBot-ThTra has a significant advantage in generating more supportive and adaptive responses.