Parallel Discourse Annotations on a Corpus of Short Texts

Parallel Discourse Annotations on a Corpus of Short Texts Manfred Stede∗, Stergos Afantenos†, Andreas Peldszus∗, Nicholas Asher†, J´er´emy Perret† ∗ Applied Computational Linguistics † IRIT CNRS University of Potsdam (Germany) Universit´e Paul Sabatier, Toulouse (France) {stede|peldszus}@uni-potsdam.de {stergos.afantenos|nicholas.asher|jeremy.perret}@irit.fr Abstract We present the first corpus of texts annotated with two alternative approaches to discourse structure, Rhetorical Structure Theory (Mann and Thompson, 1988) and Segmented Discourse Representation Theory (Asher and Lascarides, 2003). 112 short argumentative texts have been analyzed according to these two theories. Furthermore, in previous work, the same texts have already been annotated for their argumentation structure, according to the scheme of Peldszus and Stede (2013). This corpus therefore enables studies of correlations between the two accounts of discourse structure, and between discourse and argumentation. We converted the three annotation formats to a common dependency tree format that enables to compare the structures, and we describe some initial findings. Keywords: Discourse structure, Argumentation, Multi-layer annotation 1. Introduction In recent years, three approaches to analyzing and repre- senting discourse structure have resulted in various anno- tated corpora and in implemented discourse parsers: • The Penn Discourse Treebank (PDTB) annotates indi- vidual connectives with their coherence relations and their argument spans (Prasad et al., 2008). • Rhetorical Structure Theory (RST) predicts tree struc- tures on the grounds of underlying coherence relations that are mostly defined in terms of speaker intentions (Mann and Thompson, 1988). • Segmented Discourse Representation Theory (SDRT) exploits graphs to model discourse structures and de- fines coherence relations via their semantic effects on commitments rather than relative to speaker intentions (Asher and Lascarides, 2003; Lascarides and Asher, 2009). Of these, only RST and SDRT aim at predicting a full dis- course structure, and our concern in this paper is with these two theories. To date, it has been difficult to compare the two accounts on empirical grounds, since there were no directly-comparable parallel annotations of the same texts. To improve on this situation, we took an existing corpus of 112 short “microtexts”, which had already been anno- tated with argumentation structure, and added layers for RST and SDRT. To this end, we harmonized the underlying segmentation rules for minimal discourse units, so that the resulting structures can be compared straightforwardly. We implemented an approach to merging the annotations and report here on some initial observations on the correlations between RST, SDRT and argumentation in that corpus. In addition to comparing RST and SDRT, we foresee inter- esting applications of this kind of corpus data for purposes of argumentation mining. The correlations between dis- course structure and argumentation structure have not been studied yet in depth, and thus it is not clear whether estab- lished discourse parsing techniques (geared either toward RST or toward SDRT) can contribute to an automatic argu- mentation analysis, and if so, in what ways. In the following, we introduce our data set (Section 2.) and briefly describe the three layers of annotation (Sections 3.- 6.). Then, we explain the mapping of the layers to a com- …

Content truncated to 500 words