A Multi-layer Annotated Corpus of Argumentative Text: From Argument Schemes to Discourse Relations

A Multi-layer Annotated Corpus of Argumentative Text: From Argument Schemes to Discourse Relations Elena Musi§, Tariq Alhindi§, Manfred Stede †, Leonard Kriese †, Smaranda Muresan §, Andrea Rocci Columbia University§, Potsdam University†, Universita della Svizzera italiana 475 Riverside Drive New York (USA) §, Karl-Liebknecht-Str. 24-25 Potsdam (Germany) †, Via Buffi 13 Lugano (Switzerland) em3202/ta2509/sm761@columbia.edu §, stede@uni-potsdam.de, leonard.kriese@hotmail.de, andrea.rocci@usi.ch Abstract We present a multi-layer annotated corpus of 112 argumentative microtexts encompassing not only argument structure and discourse relations (Stede et al., 2016), but also argument schemes — the inferential relations linking premises to claims. We propose a set of guidelines for the annotation of argument schemes both for support and attack relations, and a new user-friendly annotation tool. The multi-layer annotated corpus allows us to conduct an initial study of dependencies between discourse relations (according to Rhetorical Structure Theory (Mann and Thompson, 1988)) and argument schemes. Our main contribution is that of offering the first resource for the combined study of (argumentative) discourse relations and inferential moves. Keywords: argumentation mining, argument schemes, discourse relations 1. Introduction Recent interest in Argumentation Mining (e.g., (Lippi and Torroni, 2016; Habernal and Gurevych, 2017) has brought to the fore the need for corpora annotated with argument information, which can be used as training data. Gener- ally, the automatic search for arguments encompasses the following steps (Peldszus and Stede, 2013): 1. the segmentation of texts into argumentative discourse units(ADUs); 2. the classification of the role (e.g., claim, premise) played by each ADU; 3. the analysis of the relations linking ADUs (e.g., sup- port, attack); and 4. the identification argument schemes, namely the im- plicit and explicit inferential relations within and across ADUs. Annotation efforts have so far provided corpora that focus on the first three steps, with genres ranging from persua- sive essays and scientific articles to online debates (e.g., (Kirschner et al., 2015; Ghosh et al., 2014; Stab and Gurevych, 2014; Walker et al., 2012). Step 4, which builds a bridge to reasoning, has not received nearly as much at- tention as the others. A notable exception is the Araucaria corpus (Reed and Rowe, 2004), which provides annotations based on Walton et al. (2008a) argument schemes. This an- notated corpus has led to work on automatically classifying argument schemes focusing on the five most frequent ones (Feng and Hirst, 2011). Other projects looked at restricted subsets of argument schemes (Green, 2017; Schneider et al., 2013) and have not yet led to publicly-available data. In this paper, we report on an annotation project that adds information about inferential rules, in the shape of argu- ment schemes, to an existing corpus that already holds an- notations of argumentation structure as well as discourse structure based on Rhetorical Structure Theory (RST) and Segmented Discourse Representation Theory(SDRT) (Stede et al., 2016). Our emphasis thus is on a multi-layer resource that allows for correlating different levels and for studying dependencies between discourse relations and ar- gument structure, or — enabled by our new result — dis- course relations and underlying …

Content truncated to 500 words