Classifying heuristic textual practices in academic discourse

In this paper, we investigate how deep learning techniques can be applied to discourse pragmatics. As a testcase we analyse heuristic textual practices, defined as linguistic implementations of decision routines in research processes in academic discourse. We develop a complex annotation scheme of pragmalinguistic categories on different levels of granularity and manually annotate a corpus of texts across various scientific disciplines. This is the basis for training recurrent neural networks to classify heuristic textual practices. Our experiments show that the annotation categories are robust enough to be recognised by our models which learn similarities of the sentence-surfaces represented as word embeddings. Our study aims at an iterative human-in-the-loop process in which manual-hermeneutic and algorithmic procedures mutually advance the insight process. It underlines the fact that the interaction between manual and automated methods opens up a promising field for further research, allowing interpretative analyses of complex pragmatic phenomena in large corpora.