Computational Narrative Understanding: The Big Picture
This paper of mine was recently published as part of the “Big Picture Workshop” that was part of EMNLP 2023. It marks a continuation of earlier work and also a further effort at consolidating the various goals and tasks that are part of the larger endeavour of using natural language processing to understand human storytelling.
The graphic below pretty well captures the larger aims. These can be broken down into five key areas:
- Narrative Element Detection. This area is concerned with granular feature detection such as characters, events, point of view, dialogue, etc. This has been bread and butter NLP and is crucial for the success of CNU.
- Multi-modality. Most of CNU has been concerned with narrative as a largely textual process. But we know narratives occur across different media, from songs to movies to paintings to illustrated books to everyday conversations. A major area of future research should address the multi-modal nature of narrative and also inter-modal differences.
- Shape, Structure, Form. No matter what term we use, narratives take shape. Because they occur over time (stories take time to tell) they can be thought of as time series data. In this time series very often resides some kind of formal patterning. This is a crucial way stories communicate meaning (beyond the individual topics and themes of the story). A lot of work has emerged in this space. Our recent work is an attempt to foreground this time series framework of narrative analysis.
- Archetypes and Collective Stories. Narrative structure isn’t just determined by time, but also by conceptual structure. The hero’s journey, the Bildungsroman, the novel of marriage. These are all archetypes that govern what is told and how. Understanding the ways in which archetypes govern larger social and cultural behavior remains a really interesting area of study. Language models in particular are going to be fundamental to this research.
- Infrastructure. In addition to the refinement and replicability of models, we still need better and more data. Most of the limitations here are due to an overly aggressive copyright landscape that has only been made worse by AIs big land grab. But we need ways of accessing and studying and sharing storytelling data to further knowledge in this field. In addition to shared tasks we also need more shared datasets. This is a hard nut to crack but essential to progress.
I hope this piece and this brief review inspire you to take on research into a particular aspect of computational narrative understanding. It’s a growing field with a lot of applications across disciplines (stories matter to a lot of fields!!). There is a ton to learn and find out and a lot of room for refining models and exploring new data. I hope this piece serves as an inspiration and acceleration for future work.