Narratives are central to how humans communicate, reason, and make sense of their experiences. They are a rich source of day-to-day knowledge and contain many social and moral norms. Therefore, narratives can be used to instill human-like communication, commonsense knowledge, and reasoning capabilities in machines. Towards this goal, Artificial Intelligence (AI) and Natural Language Processing (NLP) community, in particular, have long been interested in understanding and generating narratives. While understanding and generating narratives are natural for most people, it remains an elusive goal for AI and NLP systems, in part because it requires an understanding of key narrative elements and how they evolve and drive the story forward.
In this dissertation, we approach these challenges by designing new tasks as well as resources specifically aimed toward advancing narrative understanding and generation models. We propose to model and incorporate narrative elements that contribute to a good story, such as plot, characters, and emotions. We also introduce new modeling frameworks to incorporate user specifications in the form of narrative elements into the generation process.
First, we model the plot of the story via user-provided cue phrases as a narrative element in an interactive generation framework. We present two content-inducing approaches based on the Transformer Network for interactively incorporating cue phrases when automatically generating stories. Next, we present to model the evolution of emotions in neural story generation. We develop methods to generate stories that adhere to desired emotion arcs. We specifically designed two emotion-consistency rewards in a Reinforcement Learning framework to regularize the story generation. We then study character(s) as narrative elements. We introduce a new dataset, LiSCU, and tasks for assessing machines' understanding of characters which is of utmost importance for understanding narratives. Lastly, we focus on entities (and not just characters) as a narrative element. We propose a real-world application of generating narratives about entities grounded in world knowledge. Towards this, we present a new dataset, ranking-based generation models, and an evaluation metric of factuality.