There’s more to alternations than the main diagonal of a 2×2 confusion matrix: Improvements of MuPDAR and other classificatory alternation studies

Abstract

Corpus-based studies of learner language and (especially) English varieties have become more quantitative in nature and increasingly use regression-based methods and classifiers such as classification trees, random forests, etc. One recent development more widely used is the MuPDAR (Multifactorial Prediction and Deviation Analysis using Regressions) approach of Gries and Deshors (2014) and Gries and Adelman (2014). This approach attempts to improve on traditional regression- or tree-based approaches by, firstly, training a model on the reference speakers (often native speakers (NS) in learner corpus studies or British English speakers in variety studies), then, secondly, using this model to predict what such a reference speaker would produce in the situation the target speaker is in (often non-native speakers (NNS) or indigenized-variety speakers). Crucially, the third step then consists of determining whether the target speakers made a canonical choice or not and explore that variability with a second regression model or classifier.Both regression-based modeling in general and MuPDAR in particular have led to many interesting results, but we want to propose two changes in perspective on the results they produce. First, we want to focus attention on the middle ground of the prediction space, i.e. the predictions of a regression/classifier that, essentially, are made non-confidently and translate into a statement such as ‘in this context, both/all alternants would be fine’. Second, we want to make a plug for a greater attention to misclassifications/-predictions and propose a method to identify those as well as discuss what we can learn from studying them. We exemplify our two suggestions based on a brief case study, namely the dative alternation in native and learner corpus data.