Constructions: a New Unit of Analysis for Corpus-based Discourse Analysis

We propose and assess the novel idea of using automatically induced constructions as a unit of analysis for corpus-based discourse analysis. Automated techniques are needed in order to elucidate important characteristics of corpora for social science research into topics, framing and argument structures. Compared with current techniques (keywords, n-grams, and collocations), constructions capture more linguistic patterning, including some grammatical phenomena. Recent advances in natural language processing mean that it is now feasible to automatically induce some constructions from large unannotated corpora. In order to assess how well constructions characterise the content of a corpus and how well they elucidate interesting aspects of different discourses, we analysed a corpus of climate change blogs. The utility of constructions for corpus-based discourse analysis was compared qualitatively with keywords, n-grams and collocations. We found that the unusually frequent constructions gave interesting and different insights into the content of the discourses and enabled better comparison of sub-corpora.