Measuring the degree of transparency of English derivational suffixes

Verginica Barbu Mititelu, Gianina Iordăchioaia, Svetlozara Leseva, Ivelina Stoyanova

Abstract: We evaluate the degree of transparency of zero and overt suffixes that form verbs from nouns (e.g. -ise, -ify, ZeroV) and nouns from verbs (e.g. -er, -ion, -ment, ZeroN, a.o.) in English. These suffixes are known to be highly polyfunctional and often too ambiguous and, implicitly, non-transparent for interpretation. We build our experiment on a rich collection of 21,820 pairs of base-derived word senses extracted from the Princeton WordNet, which are annotated with a set of 14 morphosemantic relations (such as Agent, Instrument, Event, etc). We propose a machine learning method that uses the suffix and the semantic classes to automatically predict a morphosemantic relation between derivationally-related noun-verb sense pairs. The predictability of the morphosemantic relation via the machine learning mechanisms is used as a proxy for the semantic transparency of derivations (and the suffixes they involve). These results allow us to critically evaluate previous claims in the theoretical literature. We can, for instance, confirm previous observations according to which nominalizing suffixes are more transparent than the verbalizing ones, but our results challenge the claim that zero is by default less transparent than overt suffixes.

External IDs:doi:10.1007/s11525-026-09460-x