Argument Mining in Scientific Communication: Comparative Study
摘要
The paper presents the results of a study of argumentative relations based on marked-up corpora of texts of scientific communication. The authors investigated individual genre categories to assess the influence of genre features on the quality of argumentation extraction. The length of analyzed texts is at least several paragraphs, which allows quantitative and qualitative estimation of long-range relations. Long-range relations are considered to be argumentative relations between statements that are located at a distance of more than one sentence. The authors carried out a statistical comparative analysis of the relative position and distance between the components of arguments in sub-corpora that contain texts of 5 scientific and popular science subgenres. The experimental study considered the task of argument relation prediction, including long-range relations. The complexity of this task is due to several problems: issues of determining the boundaries of statements – candidates for the roles of premises or conclusions of arguments; uncertainty about the mechanism for the prediction of relations between segments that are located at long distances, and lack of agreement between experts when marking long-range relations. The paper proposes a neural network approach based on the Longformer model, which allows taking into account a fairly wide context of statements. Three models that take into account different context lengths are considered. The best results (0.72 f1-measure) were achieved on the Longformer-large model, which takes into account the context of paragraph-sized statements.