Knowledge Interpolated Conditional Variational Auto-Encoder for Knowledge Grounded Dialogues

In the Knowledge Grounded Dialogue (KGD) generation, explicit modeling of instance-variety knowledge specificity and its seamless fusion with dialogue context remains challenging. This paper presents an innovative approach, Interpolated conditional Variational auto-encoder (KIV), to address these issues. particular, KIV introduces a novel interpolation mechanism fuse two latent variables: independently encoding grounded knowledge. distinct in semantic space enables interpolated variable guide decoder toward generating more contextually rich engaging responses. We further explore deterministic probabilistic methodologies ascertain weight, capturing level specificity. Comprehensive empirical analysis conducted on Wizard-of-Wikipedia Holl-E datasets verifies that responses generated by our model performs better than strong baselines, notable performance improvements observed both automatic metrics manual evaluation.