Comparing Baselines for Corpus Analysis
Sean Wallis, Seth Mehl · Cambridge University Press eBooks · 2022
The authors review different baselines for the study of alternant choices, emphasizing that normalization to a standard number of words – while straightforward in its application – will in many cases not provide a meaningful measure of frequency. Instead, it is argued, we need a baseline indicating opportunities of use, such as phrase or sentence counts. Exemplifying their proposal with reference to get- and be-passives and the presence or absence of agentive by-phrases, the authors demonstrate a sequence of measures taken to make the quantities that are compared more meaningful and defensible, based on linguistically informed selections of baseline quantities (number of main verbs, passives or potentially alternating passives). Crucially, this process must involve a categorization of observations by the researcher to ensure that mutual substitution is plausible in each case. To calibrate this manual data verification exercise to a manageable level, the authors apply a method of uneven category sub-sampling to the data, and use it to adjust variance estimates and confidence intervals in their analysis.