Enrico Pasini2 [0000− 0002− 4525− 187X], and Daniele P.

摘要: The research of arguments in student essays has long been the subject of automatic approaches to argument mining. The task has been mostly modeled as a sequence tagging problem, where the text is either analyzed in its entirety or split into smaller homogeneous units, such as sentences or paragraphs. However, previous research has highlighted how the various essay sections may fulfill different functions, and thereby how the position of specific argument components obeys precise structural dependency criteria. Based on such underpinning we propose an approach that exploits such structural information: in this work we present a hybrid training approach that takes into account the specific structural components of the essays, in order to be able to mine different types of argument components at different levels. Our hybrid approach achieves an improvement over essay-level and paragraph-level training, in particular in the extraction of some specific argument components.