Generating hypotheses for alternations at low and intermediate levels of schematicity. The use of Memory-based Learning

Abstract According to usage-based linguistics, language variation addresses a functional need of the language user. That functional need may be dependent on the lexical realization of the varying constructions. For instance, while it may be useful to have an argument structure alternation express a particular semantic distinction for particular verbs or themes, that same distinction may be less relevant for other verbs or themes. As such, it has been argued that language variation should be investigated at low levels of schematicity, e.g. by studying argument structure alternations separately for various verbs, themes, etc. In this paper, we develop a data-driven procedure to do so, based on Memory-based Learning (MBL). The procedure focusses on generating hypotheses, is scalable, and can work with small datasets. It consists of three steps: (i) choosing features for the MBL classifier, (ii) running MBL analyses and selecting which analyses to put under further scrutiny, and (iii) inspecting which features were most useful in predicting the choice of variant in these analyses. Finally, the hypotheses that are inferred from these features are put to the test on separate data. As an example study, we investigate the Dutch naar-alternation.