Retrieval, Selection and Writing: A Three-Stage Knowledge Grounded Storytelling Model

Abstract

Storytelling is a knowledge-driven task, which requires the model to associate relevant information given a context and organize them into a reasonable story. Some knowledge-enhanced storytelling models are proposed. However, there exist certain drawbacks in them, including needing better retrieval and selection strategy. Target on these, we propose a three-stage knowledge grounded storytelling model, which combines semantic knowledge retrieval, a knowledge selection and a story generation module. We build in-domain and open-domain knowledge, which is suitable for story generation. The knowledge selection and story generation modules are built in the readily available transformer-based framework. We devise two approaches to train the knowledge selection module. One is to leverage constructed pseudo labels. Another is to jointly train the knowledge selection and story generation modules, so as to leverage the supervision of the ground-truth story. We conduct experiments on the public ROCStories dataset, and the automatic and human evaluation demonstrates the effectiveness of our method.