• LucidDream: Dynamic Story Generation through Directed Chatbot Interactions

      Stonebraker, Ryan; Metzgar, Jonathan; Lawlor, Orion; Hartman, Chris (2020-05)
      Natural Language Understanding and Generation are both areas of active research with widespread potential for story telling. This paper proposes an architecture for dynamically generating stories that allows a scene to be constructed and then dynamically written through the interaction of individual chatbots. Each chatbot in this environment is meant to mimic either the specific emotional profile of a character or holistically represent all of the character’s attributes. Chatbots are created using the conversation history so that they can understand context, a relevant sentence suggestion provided by a question-answering model to keep generated output on topic, and a finetuned version of the GPT-2 transformer-based language model to combine all of this information and generate text. This architecture serves as an ensemble method of approaching character modeling and also introduces the little-explored concept of emotional style transferring as a method for merging a story character’s emotional attributes with an independent training corpus. The question-answering model used in this study achieved 65.24% accuracy when tested on the Stanford Question-Answering Dataset and the emotion classification model achieved 57.3% accuracy on the International Survey on Emotion Antecedents and Reactions dataset. While neither of these performances are SOTA for their respective individual tasks, they are used in combination to produce state of the art directed story generation and pave the way for future research.