The Methodius Corpus of Rhetorical Discourse Structures and Generated Texts

The Methodius Corpus of Rhetorical Discourse Structures and Generated Texts Amy Isard School of Informatics, University of Edinburgh 10 Crichton Street, Edinburgh EH8 9AB amy.isard@ed.ac.uk Abstract Using the Methodius Natural Language Generation (NLG) System, we have created a corpus which consists of a collection of generated texts which describe ancient Greek artefacts. Each text is linked to two representations created as part of the NLG process. The first is a content plan, which uses rhetorical relations to describe the high-level discourse structure of the text, and the second is a logical form describing the syntactic structure, which is sent to the OpenCCG surface realization module to produce the final text output. In recent work, White and Howcroft (2015) have used the SPaRKy restaurant corpus, which contains similar combination of texts and representations, for their research on the induction of rules for the combination of clauses. In the first instance this corpus will be used to test their algorithms on an additional domain, and extend their work to include the learning of referring expression generation rules. As far as we know, the SPaRKy restaurant corpus is the only existing corpus of this type, and we hope that the creation of this new corpus in a different domain will provide a useful resource to the Natural Language Generation community. Keywords: Corpus, Natural Language Generation, Discourse Structure 1. Introduction Most rule-based Natural Language Generation (NLG) sys- tems use a pipeline of hand-crafted components to gener- ate texts (Reiter and Dale, 2000), requiring expert knowl- edge (see Section 2). There has been recent progress on using machine learning for content planning (Barzilay and Lapata, 2008; Duboue and McKeown, 2001), broad cov- erage surface realization (Rajkumar and White, 2014), and concept-to-text systems (Konstas and Lapata, 2013; Kon- stas, 2014), but these do not produce complicated discourse relations. Angeli et al. (2010) and Konstas and Lapata (2013) have researched machine learning methods to cre- ate end-to-end NLG systems, but these create more limited discourse structures than those produced by the traditional pipeline systems. To begin to fill this gap, recent research (White and Howcroft, 2015) has attempted to learn rules for the combination of clauses, using the SPaRKy Restau- rant Corpus (Walker et al., 2004; Walker et al., 2007). SPaRKy (Sentence Planning with Rhetorical Knowledge) is a sentence planner which uses rhetorical relations and adapts to a user’s individual sentence planning preferences. The SPaRKy Restaurant Corpus pairs text plans which con- tain rhetorical structure with the resulting restaurant rec- ommendation texts. White and Howcroft (2015) obtain OpenCCG logical forms (see Section 2.3) by parsing the texts, and their rule-induction algorithm learns aggrega- tion and discourse connective rules. These automatically- derived discourse rules could be used in an NLG system in place of the hand-written rules which normally form part of the generation process. White and Howcroft (2015, p.29) note that “To our knowl- edge, the SRC remains the only publicly available corpus of input-output pairs for an NLG system using discourse structures with rhetorical relations.” We …

Content truncated to 500 words