A regression model for the English benefactive alternation : An efficient, practical, actually usable approach

Abstract

In this paper, we use logistic regression modelling to predict the English benefactive alternation
\n(He baked me a cake vs. He baked a cake for me). We developed a data set
\nconsisting of 107 instances in adult writing and 36 in the writing of 8-to-12-year-olds, and
\nannotated them with 13 syntactic, semantic and discourse features. We show that a model
\ntrained and tested on the adult data reaches a prediction accuracy of 86.9%. Due to the
\nsmall number of data instances, our model includes only 4 significant effects and shows
\nconsiderable overfit (reaching 79.6% accuracy in a ten-fold cross-validation setting). The
\nregression coefficients found are similar to those found in the model for the to-dative alternation
\n(Bresnan et al. 2007). When applying the adult model to the instances in child
\nwriting, 80.6% is predicted correctly. We conclude that there are no indications of major
\ndifferences either between the to-dative and benefactive alternation in adult language, nor
\nbetween the benefactive alternation in adult language and that in child language.