The Exploitation of Spatial Information in Narrative Discourse

We present the results of several machine learning tasks that exploit explicit spatial language to classify rhetorical relations and the spatial information of narrative events. Three corpora are annotated with figure and ground (granularity) relationships, mereotopologically classified verbs and prepositions, and frames of reference. For rhetorical relations, Naı̈ve Bayesian models achieve 84.90% and 57.87% accuracy in classifying NARRATION and BACKGROUND / ELABORATION relations respectively (16% and 23% above baseline). For the spatial information of narrative events, K* models achieve 55.68% average accuracy (12% above baseline) for all spatial information types. This result is boosted to 71.85% (28% above baseline) when inertial spatial reference and text sequence information are considered. Overall, spatial information is shown to be central to narrative discourse structure and prediction tasks.