A Flexible NLP Pipeline for Computational Narratology

Temporal dependencies reveal interesting insights into the semantic discourse structure of narrative texts. The investigations of literary scientists are, as of today, mostly based on labor-intensive manual annotations. Computational Narratology, an important subtopic of the Digital Humanities, aims at facilitating annotations and supporting literary scientists with their analyses. According to Mani (2013), one aspect of Computational Narratology focuses on exploring and testing literary hypotheses through mining narrative structures from corpora. In the context of the BMBF-funded eHumanities project heureCLÉA, we address temporal phenomena in literary text, a genre whose temporal phenomena are different from others. For example, it is often not possible to anchor temporal expressions to real points in time, but literary texts tend to have their own time frame. Our project partners, as well as many other humanists, use CATMA, a comprehensive graphical tool for annotating data. Interfacing NLP with CATMA could drastically reduce the effort of manual annotation. The goal of heureCLÉAis to provide users with a collaborative annotation environment for tagging temporal phenomena in documents, with simple annotations (e.g., temporal expressions) being added automatically, and more complex annotations (e.g., time shifts and ellipses) being suggested. Users can correct automatic annotations, and user feedback will be used to apply machine learning techniques to improve future annotation suggestions. In the following, we outline our flexible architecture for NLP in the domain of narrative texts, as well as promising first results for annotating the tense of sub-sentences to demonstrate the effectiveness of our approach.