Computer-based Analysis of Argumentation Patterns and Strategies in Scientific Communication Texts

The article presents the results of analyzing argumentation patterns that are employed in scientific communication texts. We define argumentation patterns as models of separate arguments and their composite structures repeated across corpus texts. The corpus consists of 98 texts in three different genres: 50 scientific articles from the scientific electronic library CyberLeninka (https://cyberleninka.ru/), 20 popular science articles from the forum habr.com/ru (hereafter ‒ Habr articles), 28 scientific news texts from the Poisk aggregator of scientific news (poisknews.ru). Five experts in theoretical and applied linguistics have performed double annotation of the corpus texts using ArgNetBank Studio platform. Calculated inter-annotator agreement scores correspond to the “substantial agreement” range. Argumentation annotation follows the Argument Interchange Format standard. The results of annotation and the object of the analysis are argumentation graphs with two node types: information nodes (which correspond to argumentative statements) and scheme nodes (which, for each link between statements, indicate its reasoning model in accordance with Walton’s compendium). The study addresses the graphs obtained from the abovementioned graphs by excluding all information nodes and merging the incoming and outcoming links of each excluded information node into one link (thus the obtained graphs consist exclusively of scheme nodes) and the subgraphs of these obtained graphs. We perform a frequency-based analysis of subgraphs by the Frequent Subgraph Mining Method combined with Cordella VF2 algorithm implementation from the NetworkX library for isomorphism check. The analysis results in identification of argumentation patterns (repeating subgraphs from 1 to 9 node in size) that are employed in texts of all genres, one specific genre, as well as patterns repeated within separate texts. We show that scientific articles of the corpus exhibit the greatest stability and diversity of argumentation patterns, while popular science texts are marked by the active use of polemic argumentation, and patterns in scientific news are based on two reasoning models characteristic of compact argumentation. Patterns that repeat in separate texts show the same tendencies as genre-specific patterns. Using the identified patterns, we conduct clustering of the corpus texts with Ward and K-means algorithms. Resulting clusters correspond to groups of texts with similar argumentation strategies and exhibit homogeneity of texts by genre and thematic field. Identified patterns can be used not only for classifications but also for assessing the argumentative organization of texts extracting persuasive argumentation, its generation etc. Works on this topic are still underrepresented for texts in English, while for texts in Russian, such automatic analysis of argumentation patterns has not yet been performed.