Intertextual Correspondence for Integrating Corpora

Intertextual Correspondence for Integrating Corpora Jacky Visser, Rory Duthie, John Lawrence, Chris Reed Centre for Argument Technology University of Dundee, Dundee DD1 4HN, United Kingdom {j.visser, r.w.duthie, j.lawrence, c.a.reed}@dundee.ac.uk Abstract We present intertextual correspondence (ITC) as an integrative technique for combining annotated text corpora. The topical correspon- dence between different texts can be exploited to establish new annotation connections between existing corpora. Although the general idea should not be restricted to one particular theoretical framework, we explain how the annotation of intertextual correspondence works for two corpora annotated with argumentative notions on the basis of Inference Anchoring Theory. The annotated corpora we take as examples are topically and temporally related: the first corpus comprises television debates leading up to the 2016 presidential elections in the United States, the second corpus consists of commentary on and discussion of those debates on the social media platform Reddit. The integrative combination enriches the existing corpora in terms of the argumentative density, conceived of as the number of inference, conflict and rephrase relations relative to the word count of the (sub-)corpus. ITC also affects the global properties of the corpus, such as the most divisive issue. Moreover, the ability to extend existing corpora whilst maintaining the level of internal cohesion is beneficial to the use of the integrated corpus as resource for text and argument mining based on machine learning. Keywords: argument, corpus, debate, dialogue, intertextuality, Reddit, US presidential elections 1. Introduction In many regards, a larger corpus is a better corpus (all other things being equal). Additionally, for some purposes, a diverse corpus covering various communicative genres or data from different modalities can be of great value. In a study of argumentative conduct in television debates leading up to the 2016 presidential elections in the United States, we are looking for both. We need a large corpus for the purpose of training machine learning algorithms for argument mining (Palau and Moens, 2009; Feng and Hirst, 2011; Stab and Gurevych, 2014) – with a size of 97,999 words (tokens), our US2016 corpus is the largest corpus of dialogical argumentation annotated for both inferential and discursive structure. We also want the corpus to be amenable to automated comparison between the character- istics of argumentative discussions in different communica- tive genres, as a means of quantifying the uptake certain topics from the television debates find in online social me- dia reactions. To construct a large corpus, it might be sufficient to collate various (preferably consistently annotated) corpora into one super-corpus. However, if any of the studies for which the text corpus is used relies upon the relationships between different parts of the corpus, then mere collation will not suffice. To properly integrate corpora, we propose that in- tertextual correspondence can be used as a method to pre- serve dialogical cohesion by functionally connecting con- tent from two or more corpora on the basis of their topical correspondence. In the present paper, we introduce the notion of ‘intertextual correspondence’ (ITC) in relation to existing work (Sec- tion …

Content truncated to 500 words