Simple Contrastive Learning with Knowledge Graphs for Story Generation

Abstract

Story generation stands as a crucial, yet formidable task, necessitating a profound grasp of subtle, often unspoken knowledge, along with context-specific cues to craft compelling narratives. The core challenges involve effectively harnessing this implicit knowledge and augmenting narrative diversity. Addressing these hurdles, this paper introduces SimCoS - an acronym for Simple Contrastive learning with knowledge graphs for Story generation. SimCoS significantly bolsters language models’ performance in story generation by sharpening their ability to perceive and discern input nuances, aided by a specialized story knowledge graph. Moreover, SimCoS debuts a novel methodology for evaluating ‘negative instances’ using subgraphs, which enables it to calculate logical distances between inputs, thereby enhancing differentiation among diverse stories. Through both automatic and manual assessments, SimCoS has been shown to outperform leading counterparts in two story generation tasks, marking improvements of 13.4% and 4.5%, respectively. This advancement not only yields narratives of superior quality and logical coherence but also elevates their diversity.