In the biomedical domain, authors publish their experiments and findings using a quasi-standard coarse-grained discourse structure, which starts with an introduction that sets up the motivation, continues with a description of the materials and methods, and concludes with results and discussions. Over the course of the years, there has been a fair amount of research done in the area of scientific discourse analysis, with a focus on performing automatic recognition of scientific artefacts/conceptualisation zones from the raw content of scientific publications. Most of the existing approaches use Machine Learning techniques to perform classification based on features that rely on the shallow structure of the sentence tokens, or sentences as a whole, in addition to corpus-driven statistics. In this article, we investigate the role carried by the deep (dependency) structure of the sentences in describing their rhetorical nature. Using association rule mining techniques, we study the presence of dependency structure patterns in the context of a given rhetorical type, the use of these patterns in exploring differences in structure between the rhetorical types, and their ability to discriminate between the different rhetorical types. Our final goal is to provide a series of insights that can be used to complement existing classification approaches. Experimental results show that, in particular in the context of a fine-grained multi-class classification context, the association rules emerged from the dependency structure are not able to produce uniform classification results. However, they can be used to derive discriminative pair-wise classification mechanisms, in particular for some of the most ambiguous types.
展开▼
