Abstract
In this paper, we apply distributional methods to Czech data to compare the predictions of two views of inflectional paradigms, as systems of orthogonal morphosyntactic feature oppositions, or as systems of multilateral contrasts between pairs of morphologically related words, not necessarily reducible to orthogonal features.
We define two predictive tasks that probe what it means for two pairs of paradigm cells to contrast in the same features: in the first, we train a classifier to discriminate between two paradigm cells; in the second, we train a family of models to predict the vector of the word in one cell from that of the word in another cell. By varying the choice of training and test data, we show that (i) a model trained on data that contrast in a manner orthogonal to its test data performs on average at chance level, while (ii) a model trained on data that contrast in a manner parallel to its test data performs on average better than chance but still worse than a model trained on the same pair of cell used for testing. This is incompatible with the predictions of a reductive view of paradigms as systems of feature contrasts.