Multi-lingual Argumentative Corpora in English, Turkish, Greek, Albanian, Croatian, Serbian, Macedonian, Bulgarian, Romanian and Arabic

Multi-lingual Argumentative Corpora in English, Turkish, Greek, Albanian, Croatian, Serbian, Macedonian, Bulgarian, Romanian and Arabic Alfred Sliwa, Yuan Ma, Ruishen Liu, Niravkumar Borad Seyedeh Fatemeh Ziyaei, Mina Ghobadi, Firas Sabbah, Ahmet Aker Department of Computer Science and Applied Cognitive Science University of Duisburg-Essen Lotharstraße 65, 47057 Duisburg {alfred.sliwa.92, yuan.ma, ruishen.liu, niravkumar.borad seyedeh.ziyaei, mina.ghobadi, firas.sabbah}@stud.uni-due.de a.aker@is.inf.uni-due.de Abstract Argumentative corpora are costly to create and are available in only few languages with English dominating the area. In this paper we release the first publicly available corpora in all Balkan languages and Arabic. The corpora are obtained by using parallel corpora where the source language is English and target language is either a Balkan language or Arabic. We use 8 different argument mining classifiers trained for English, apply them all on the source language and project the decision made by the classifiers to the target language. We assess the performance of the classifiers on a manually annotated news corpus. Our results show when at least 3 to 6 classifiers are used to judge a piece of text as argumentative an F1-score above 90% is obtained. Keywords: Multilingual Argument Annotations, Argument Mining, Parallel Corpora 1. Introduction Argument mining refers to the automatic extraction of argu- ments from natural texts. An argument consists of a claim (also referred to as the conclusion of the argument) and sev- eral pieces of evidence called premises that support or re- ject the claim (Lippi and Torroni, 2016b). Identifying argu- ments in large volumes of textual data has the potential to revolutionarise our access to information. Argument based search for information would for example facilitate indi- vidual and organisational decision-making, make learning more efficient, enable quicker reporting on present and past events, to name just a few broad applications. Even more important is argument mining in the multi-lingual context, by which argument based retrieval would be available to people in the language of their preference. Current studies report methods for argument mining in legal documents (Reed et al., 2008), persuasive essays (Nguyen and Litman, 2015), Wikipedia articles (Levy et al., 2014; Rinott et al., 2015), discussion fora (Swanson et al., 2015), political debates (Lippi and Torroni, 2016a) and news (Sardianos et al., 2015; Al-Khatib et al., 2016). In terms of methodology, supervised machine learning is a central technique used in all these studies. This assumes the availability of data sets – argumentative texts – to train and test the argument mining models. Such data sets are readily available in English and – although in comparably smaller quantities – in very few European languages such as German or Italian. Languages other than these are cur- rently neglected. Due to this lack of data the research and development of argumentation mining outside English and few European languages is very limited, rendering multi- lingual argument mining and language independent argu- ment based retrieval impossible. In this research we aim to fill this gap. We aim to create multi-lingual corpora annotated with argumentative struc- tures automatically. For this purpose we make use of …

Content truncated to 500 words