Towards Empathy Learning, Socially-Aware Agents

Towards Empathy Learning, Socially-Aware Agents

We introduce a novel, model-agnostic, and dataset-agnostic method that approximates interactive human evaluation in the open-domain dialog. We develop an off-policy reinforcement learning (RL) scenario and show that solely relying on explicit human preferences is not as effective as training with implicit human rewards. We build a novel hierarchical RL model and demonstrate its effectiveness in reducing repetitiveness or toxicity.

Publications: NeurIPS’19, EMNLP’20, AAAI’20, NeurIPS’19 Conv. AI workshop</p>

Talks: NeurIPS'19 WiML workshop, NeurIPS'19 Conv. AI workshop

Awards: MIT Quest for Intelligence, MIT Stephen A. Schwarzman College of Computing, Machine Learning Across Disciplines Challenge

More: Deployment Website, Modeling code, Server code