Computational Models for Story Plot Generation and Summarisation

Abstract

Story plot generation and summarisation are two open problems in natural language generation and understanding. Initial approaches to story generation and summarisation relied heavily on extensive knowledge engineering. To overcome this expensive dependency, many researchers have designed neural-based approaches. However, the current neural models have their fair share of weaknesses. Their generated story lacks many characteristics like a central theme, purpose, etcetera. It would hardly engage any reader. Additionally, these models are very opaque, hardly shed any light on the storytelling process, and are far from explainable AI. This thesis proposes a mathematical model of story plot generation and summarization. We hypothesise that a story is a series of states which requires certain conditions to be met before the transition of the states. Therefore, we use Petri net as a causal backbone of the model. We populate its knowledge from open data sources like ConceptNet and use a transformer model like Comet to predict the knowledge for unknown cases. This way, we synergise the best of both worlds, where we have transparency and control like conventional models, while we do not have to engineer knowledge like the neural models manually. In addition to proposing the model, we formalise many aspects of storytelling, like consistency and coherence. We propose multiple approaches utilising the model and formalisation and address many challenges in the literature, like avoiding repetitive stories and handling flat characters. At last, we evaluate the proposed approaches and present our findings.