Knowledge Bridging for Empathetic Dialogue Generation

Lack of external knowledge makes empathetic dialogue systems difficult to perceive implicit emotions and learn emotional interactions from limited history. To address the above problems, we propose leverage knowledge, including commonsense lexical explicitly understand express in generation. We first enrich history by jointly interacting with construct an context graph. Then representations knowledge-enriched graph distill signals, which are prerequisites predicate expressed responses. Finally, generate response, cross-attention mechanism dependencies Extensive experiments conducted on a benchmark dataset verify effectiveness proposed method. In addition, find performance our method can be further improved integrating pre-trained model that works orthogonally.