The present paper, in the domain of theoretical and cognitive linguistics, aims to build on previous work by authors like Aijmer (1997, 2015), Kaltenböck (2009, 2010) or Simon-Vandenbergen (2000) focussing on the sequence "I think" as a discourse marker, used to indicate stance or epistemic modality. It is well known that "I think" is liable to assume a variety of different values, which Kaltenböck (2010), for example, identifies as "shielding", "approximator", "structural" or "booster" functions. We hypothesise that "I think" is not inherently ambiguous, but that the various values reflect specific configurations of "I think", and that these are dependent on identifiable contextual features. The present study aims to explore this hypothesis, firstly, by a corpus-based investigation of collocational affinities of the sequence, which will reveal a number of characteristic environments. This part of the study will use the British National Corpus, accessed via the BNCweb interface and queried for specific collocations. Secondly, we will endeavour to elaborate an enunciative description of "I think" in terms of a basic schematic form, which undergoes certain controlled and calculable deformations to generate local "shapes", i.e. contextually situated values (Culioli 1990).
The theoretical objectives of the present contribution are to show 1) that if a marker is not contextually situated then it can only be described in terms of its potential for meaning; 2) that the value of a marker depends both on the surrounding context (including prosody), and also on its own specific latitudes of deformability; 3) that such hypotheses, which are recognisably associated with enunciative approaches to languages (and qualitative approaches more generally) can be refined and strengthened with the aid of targeted corpus queries and the resulting quantitative analysis.