Productivity-Based Analysis of Argumentation Patterns across Texts of Different Genres

The presented work describes the method of the automatic identification of argumentation patterns and the results of their analysis in terms of usage and productivity. The employed corpus contains a total of 4328 arguments extracted from manually constructed argumentative annotations for texts from three genres of scientific communication: short scientific papers (50 texts), scientific news (28), analytic online articles in popular science style from the Habr collaborative blog (20). Argumentative annotations of these texts, constructed in accordance with the Argument Interchange Format standard, take form of oriented graphs with nodes of two types: information nodes (statements) and scheme nodes connecting them (reasoning models from Walton’s compendium). We define argumentation patterns as subgraphs with three and more connected arguments, represented at the abstract level of reasoning models, that repeatedly occur across texts by different authors. As argumentation graphs do not explicitly mark complex argumentation patterns, their identification relies on the methods of the Frequent Subgraph Mining. The comparison of subgraphs is performed with the use of the VF2 algorithm for the exact isomorphism check, implemented in the NetworkX graph library. The analysis of the identified patterns reveals, one, the dependence of argumentation schemes productivity in forming patterns on their semantic properties rather than frequencies of standalone usage, and two, the absence of a clear correlation between the patterns productivity and consistency of separate schemes usage across the three genres. The specification of productive patterns at the level of root schemes facilitates the identification of schemes in machine construction of argumentative annotations.