Generative Model Using Knowledge Graph for Document-Grounded Conversations

Document-grounded conversation (DGC) is a natural language generation task to generate fluent and informative responses by leveraging dialogue history document(s). Recently, DGCs have focused on fine-tuning using pretrained models. However, these approaches problem in that they must leverage the background knowledge under capacity constraints. For example, maximum length of input limited 512 or 1024 tokens. This fatal DGC because most documents are longer than length. To address this problem, we propose document-grounded generative model graph. The proposed converts sentences extracted from given document(s) into graphs fine-tunes We validated effectiveness comparative experiment well-known Wizard-of-Wikipedia dataset. outperformed previous state-of-the-art our experiments Doc2dial