The Combined Approach to Identifying Argumentation Structures in Short Scientific Papers

The paper described the method for identifying argumentation structures in scientific texts in Russian language. This approach is aimed at automating argumentation annotation of text sets. It combines the machine learning methods and rule-based search through patterns that are based on a dictionary of argumentation markers. The combined method covers the four stages of modelling the argumentation structure of a text by an annotator: 1) identification of statements in a text, their classification into argumentative and non-argumentative; 2) detection of statements connections; 3) specification of statement roles in arguments (premises, conclusions); 4) identification of exact reasoning model in an argument (Analogy, Example, Verbal Classification). Stages 1) and 3) employ the machine learning methods (LogReg, SVM, MLP, MNB), the stage 2) uses the markers dictionary, the stage 4) relies on search patterns in form of regular expressions. The dataset for evaluating the method consists of argumentation annotations for 29 texts of short scientific papers from two thematic areas (information science and linguistics). These annotations are constructed by human experts with the ArgNetBankStudio platform and contain 1259 arguments and 1309 statements. The paper provided the quality scores for the identification of argumentation components at every stage. These scores showed that the combined method is applicable to the partial automatization of annotating argumentation.