Using machine learning models to interpret disciplinary styles of metadiscourse in dissertation abstracts

Abstract

This paper presents the results of a study of disciplinary stylistic differences among dissertation abstracts from physics, psychology, and philosophy. Based on differences in relative frequencies of metadiscourse terms as provided by Hyland (2005), we used a machine learning approach to construct SMO vector support models of each discipline whose average accuracy (88.3%) surpassed a baseline model by 22%. We found that model term weights supported the findings of previous qualitative research regarding differences between disciplines and by extension between hard sciences, social sciences, and humanities. Given the success of the metadiscourse-based model, we conclude by proposing an expanded study to investigate disciplinary style both across disciplines and over time.