Computational Analysis of Arguments and Persuasive Strategies in Political Discourse

Various persuasive strategies are employed in advancing argumentation. This dissertation presents the first computational work in analyzing persuasive strategies in monological and dialogical argumentation in natural language. I begin with reputation defence strategies and show to what extent human annotators agree on these strategies. I present the first manually annotated corpus of parliamentary debates annotated with the most agreed upon face-saving strategies and show that linguistic features automatically extracted from the text of debates can differentiate between these strategies. Having shown the effectiveness of discourse parsing features in the classification of reputation defence strategies, I hypothesize that by directly using the effective features for discourse parsing, the classification results can be improved. My experiments validate this hypothesis and show that the developed methods can automatically label speeches with these strategies. I then explore whether we can automatically predict the language of face-saving in speeches and show that by leveraging the contextual information of the speeches, we can reliably distinguish between reputation defence from non-defence. I further investigate whether we can automatically classify statements in face-threatening and face-saving speeches based on truthfulness using the effective linguistic features introduced in the prior literature and show that while some of these features help identify the expression of dodge, they are not very effective in identifying the truthfulness of the statements. I further operationalize framing analysis as a classification task and show that neural language models can capture the abstract representations of frames more effectively. My experiments also show that frames are transferable across genres. Finally, in collaboration with several researchers, we examine to what extent expert and lay annotators can evaluate argumentation aspects, and show that the agreement of both groups is limited.Ph.D