Modeling People in Story Generation
Speaker: Snigdha Chaturvedi (UNC)
Date and Time: Friday, November 18 at 3:30pm CT
Automatic story generation is the task of designing NLP systems that, given a prompt, can produce the text of a story. Most methods for this problem focus on modeling events and their coherence. An alternate perspective to story generation can be from the viewpoint of people described in the story. In this talk, I describe two aspects of modeling people in story generation -- modeling emotions and social relationships. In the first part of the talk, I describe our story generation approach to incorporate a desired emotion arc for the protagonist. We use Reinforcement Learning to encourage the generation model to adhere to the desired emotion arc and show that our approach results in stories that fit the emotion arc while maintaining their coherence. In the second part of the talk, I describe our story generation approach to incorporate a desired social network demonstrating relationships between various people to be mentioned in the story. We use a latent-variable-based model that generates the story one sentence at a time. It uses a latent variable to pick the relationship, if any, to be exhibited in the next sentence to be generated. It then generates the story sentence conditioned on the latent variable. Our experiments show that this model can generate coherent stories that can reflect the desired social network. Also, the latent variable-based design results in an explainable generation process.
Snigdha Chaturvedi is an Assistant Professor of Computer Science at the University of North Carolina, Chapel Hill. She specializes in Natural Language Processing with an emphasis on narrative-like and socially aware understanding, summarization, and generation of language. Previously, she was an Assistant Professor at UC-Santa Cruz, and a postdoctoral fellow at UIUC and UPenn working with Dan Roth. She earned her Ph.D. in Computer Science from UMD in 2016, where she was advised by Hal Daume III. Her research has been supported by NSF, Amazon, and IBM.