Indicators of Argument-conclusion Relationships. An Approach for Argumentation Mining in German Discourses

Abstract

[In this paper we present a new methodological approach for the analysis of public discourses aiming at the semi-automated identification of arguments by combining methods from discourse analysis with methods from Natural Language Processing. Discourses evolve over long periods of time and, consequently, form a broad database. Up to now, the analysis of discourses is hitherto performed primarily by hand, i.e., only small corpora or discourse fragments can be analyzed. Inevitably, this leads to lengthy and expensive annotation. Thus, there is a growing interest to overcome these methodological challenges by the use of computer-based methods and tools for the semi-automated analysis. However, there are only few approaches known that focus on the analysis of discourses and the (semi-)automated identification of arguments therein (e.g. Reed at al., 2008; Liakata et al., 2012; Ashley and Walker, 2013). Particularly, approaches that can be explicitly used for the analysis of German-language discourses exist only in initial stages. Therefore, we suggest a fine-grained semi-automated approach based on multi-level annotation that focuses on linguistic means as indicators of arguments. The aim is to identify regularities, respectively, indicators in the linguistic surface of the discourse (e.g. recurring lexical and typographical characteristics), which indicate the occurrence of certain arguments (e.g. premise). In this paper, we focus on the identification of indicators of argumentconclusion relationship: conclusive connectors or conclusiva, that are typically adverbs such as hence, consequently, therefore, thus, because (Govier, 2013; see example below): Die Campusbahn werde den Individualverkehr verdrangen, weil die Stadt eng bebaut sei. Schon in den 1970er Jahren sei deshalb das Aus fur die Strasenbahn besiegelt worden. [The campus train will displace the individual traffic because the city is densely built. Therefore, the end for the tram was sealed in the 1970s.] As an application example, a small corpus consisting of 21 newspaper articles is analyzed. The corpus belongs to the interdisciplinary project Future Mobility (FuMob), which is funded by the Excellent Initiative of the German federal and state governments. The methodological approach consists of three steps, which are performed iteratively: (1) manual discourselinguistic argumentation analysis, (2) semiautomatic Text Mining (PoS-tagging and linguistic multi-level annotation), and (3) data merge. (1) Discourse-linguistic argumentation analysis: First, the data is manually analyzed. Objectives of the analysis are (i) identifying discourserelevant arguments, (ii) forming argument classes, and (iii) determining the significance of an argument in the discourse (Niehr 2004). To determine the significance of an argument the use by various discourse participants is analyzed and quantified. The argument-use can be categorized as argumentative, positively cited, negatively cited or neutrally cited. In addition, to identify arguments and their use in public discourse, the analysis also aims to detect and characterize discourse participants who use similar arguments.