This paper addresses the challenge of dialogue agents recognizing and responding to feelings in conversations, a key skill for effective communication. The authors propose a new benchmark and dataset, EMPATHETICDIALOGUES, which contains 25,000 conversations grounded in emotional situations. They find that dialogue models trained on this dataset are perceived as more empathetic by human evaluators compared to models trained on large-scale internet conversation data. The paper also presents methods for improving empathetic responses using the dataset, including retrieval-based and generative approaches. Experiments show that incorporating ED candidates or fine-tuning models on ED data can enhance empathetic responses, with minimal additional training resources required. The work contributes to the field by providing a new benchmark and dataset for evaluating and improving empathetic dialogue generation.This paper addresses the challenge of dialogue agents recognizing and responding to feelings in conversations, a key skill for effective communication. The authors propose a new benchmark and dataset, EMPATHETICDIALOGUES, which contains 25,000 conversations grounded in emotional situations. They find that dialogue models trained on this dataset are perceived as more empathetic by human evaluators compared to models trained on large-scale internet conversation data. The paper also presents methods for improving empathetic responses using the dataset, including retrieval-based and generative approaches. Experiments show that incorporating ED candidates or fine-tuning models on ED data can enhance empathetic responses, with minimal additional training resources required. The work contributes to the field by providing a new benchmark and dataset for evaluating and improving empathetic dialogue generation.