The duration of a narrative—that is, how long the events described in the story world take to unfold—is a feature of general interest to narrative understanding, and has been a topic of interest to theoreticians of narrative for some time. We combine prior work on timeline extraction (the TLEX algorithm) with an event duration estimation method that uses temporal pattern mining from large text corpora, to demonstrate a method for estimating the duration of narratives. We first gathered over 400K event durations mined from nearly 400M words from two corpora (the iWeb corpus and LexusNexis) using 10 hand-crafted temporal patterns. We then apply our approach to 30 selected stories from two corpora that already have annotated gold standard events and times (the ProppLearner corpus and the TimeBank corpus), estimating the duration of each story. We then conducted a preliminary evaluation using human judgements, showing that the durations extracted by the system are judged reasonable by people approximately 70% (for folktales) and 28% (for news stories) of the time. The gap between actual and desired performance reveals several challenges which are of interest, related both to duration estimation and its evaluation.
