Employing Argumentation Patterns for Genre Classification of Scientific Communication Texts

The article presents the results of examining applicability of argumentation patterns to automatic genre identification for scientific communication texts. We define argumentation patterns as reasoning models (from Walton's compendium, also known as argumentation schemes) and their composite structures (in connected arguments) that repeat across different texts. The experiment employs a corpus of 98 texts: 28 scientific news, 20 popular science Habr articles, 50 scientific papers from two thematic fields (information technologies and linguistics, 25 articles for each). Altogether the corpus texts contain 5042 statements and 4328 arguments. Five experts have annotated argumentation in corpus texts in accordance with the Argument Interchange Format standard and reached agreement values from the “substantial” range. Annotation of each text consists in representation of its argumentation as a connected oriented graph. Argumentation patterns correspond to subgraphs in modified graphs formed exclusively by nodes with argumentation schemes (without statement nodes). We perform the genre classification by using three algorithms (multinomial naive Bayes, support vector machine, multilayer perceptron) on five test sets with four feature sets: elementary patterns (separate schemes), schematic bigrams (sequences of two schemes in connected arguments), complex patterns (structures of three and more schemes), and the full range of argumentation patterns (combination of three previous feature sets). The classification results demonstrate the applicability of argumentation patterns to the task of genre classification. The highest quality scores belong to the multilayer perceptron on the full range of patterns: precision, recall, and F-measure respectively equal to 0.97, 0.93, 0.95 for micro-average.