Generating Knowledge Graphs using GPT-J for Automated Story Generation Purposes

Abstract

Automated story generation is a recent challenge in artificial intelligence that aims to create stories in coherent ways. While it’s made much progress, stories are still not coherent; however, a possible solution is to create a knowledge graph, or map of the world, to guide a story. This paper aims to create a knowledge graph using a large pre-trained model called GPT-J. GPT-J was prompted based on a type of noun, and responses were then filtered and shortened to get a suitable answer to be added to the knowledge graph. GPT-J showed to contain many words that were not included in another commonly used knowledge graph called ConceptNet5. Most notably, nouns referring to people had a far lower amount of related words in ConceptNet5 compared to those produced by this system.However, there are still serious limitations of this system since human respondents only responded saying 41% of responses were good responses; thus major improvements are needed.