Salience for Narrative Similarity and Clustering

For my Ph.D. thesis I developed a distance metric for measuring story plan similarity according to a psychological model of narrative comprehension. To validate the metric I collected a human dataset, published here: Story Similarity Dataset.

Among other uses, the salience-based distance metric allows domain authoring tools to cluster stories (narrative plans) meaningfully, and provide critical feedback to domain authors about what kinds of stories are modeled by their domain.

To evaluate my proposed technique for summarizing solution spaces, I made a data collection tool that lets users configure a story world to their liking and explore it by clicking on clusters to sample random example stories.

There are three versions: The Circles and Tree versions each use a different type of clustering, and the Control version uses no clustering (all example stories are randomly sampled from the full set). You can toggle between the three different versions at the bottom of the page.

The tool simulates domain authoring by allowing users to configure the domain using a predefined set of variables. It then assesses their understanding of the domain they configured with a series of 14 questions: Given a potential story, answer whether or not that story is modeled. (This version is a demo and will not collect any data.)

Our study results demonstrated that both of the summary tools helped people answer the questions more accurately. For elaborate detail, see Chapter 3 of my dissertation.