TreeAnnotator: Versatile Visual Annotation of Hierarchical Text Relations

TREEANNOTATOR: Versatile Visual Annotation of Hierarchical Text Relations Philipp Helfrich, Elias Rieb, Giuseppe Abrami, Andy L¨ucking, Alexander Mehler Goethe University Frankfurt, Text-Technology Lab Robert-Mayer-Straße 10, 60325 Frankfurt am Main helfrich@cs.uni-frankfurt.de, s0294036@stud.uni-frankfurt.de, {abrami, luecking, mehler}@em.uni-frankfurt.de Abstract We introduce TREEANNOTATOR, a graphical tool for annotating tree-like structures, in particular structures that jointly map dependency relations and inclusion hierarchies, as used by Rhetorical Structure Theory (RST). TREEANNOTATOR is browser-based, embedded within the UIMA framework and provides two visualization modes. TREEANNOTATOR’s interoperability exceeds similar tools, providing a wider range of formats, while annotation work can be completed more quickly due to a revised input method for RST dependency relations. TREEANNOTATOR offers a multiple window view, which allows users to inspect several annotations side by side. For storing and versioning annotations, the UIMA Database Interface (UIMA DI) was developed to save documents based on a pre-defined type system. These features not only connect TREEANNOTATOR annotations to modern technological and dialog theoretical work, but set it apart from related tools. The ease of use of TREEANNOTATOR and its newly designed user interface is evaluated in a user study consisting of annotating rhetorical relations with TREEANNOTATOR and the classic RSTTool. Keywords: text annotation, text visualization, multiple views, rhetorical structure theory, UIMA, database 1. Introduction Over the past years, a lot of annotation tools have been de- veloped for various tasks (for a recent process and feature oriented overview see (Finlayson and Erjavec, 2017); for an overview of multimodal tools see (Cassidy and Schmidt, 2017)). For some reason, however, the domain of discourse annota- tion, most notably in terms of Rhetorical Structure Theory (RST) (Mann and Thompson, 1988), has not gained much attention for quite a long time. Most RST annotations, with the RST Discourse Treebank (Carlson et al., 2003)(Carlson et al., 2014) leading the way, are carried out with the orig- inal RST Annotation Tool (O’Donnell, 1997) or its exten- sion, the ISI RST Annotation Tool (Marcu, 1999). While still usable, both tools, however, are not maintained any more and do not comply to current annotation frameworks. Due to such reasons, a third, browser-based RST annotation tool has been developed recently: rstWeb (Zeldes, 2016). Why, then, is there need for a new tool like TREEANNO- TATOR? We highlight three reasons in the following: Firstly, TREEANNOTATOR, unlike rstWeb, follows the UIMA framework, which supports current requirements of an- notation pipelines (Wilcock, 2017). Note that also the Atomic/ANNIS frameworks (Druskat et al., 2014; Krause and Zeldes, 2016) are not based on UIMA, and hence TREEANNOTATOR is a module following a technologically complementary approach, more related to the “NLP-affine” text annotation tool BRAT (Stenetorp et al., 2012) (which in turn is the technological and design source for UIMA-based WebAnno (de Castilho et al., 2014)). Even extending such frameworks, we developed an UIMA database interface in order to provide a fine-grained type system-compliant stor- age of documents. Secondly, users of TREEANNOTATOR benefit from dif- ferent visualizations. The usefulness of visualizing anno- tations has been independently highlighted, for instance by (Finlayson and …

Content truncated to 500 words