[PS-3.59] Relative entropy effects in Dutch and German word reading: A discrimination learning account

van den Hoven, E. & Bradley, L.

University of Freiburg, Germany

We describe a series of morphological analyses on lexical decision and eye-tracking data from Dutch and German. We first looked for Dutch analogues of the relative entropy effect demonstrated by Milin et al (2009), whereby words associated with idiosyncratic contexts are on average recognized less easily when not presented in those contexts (e.g. 'oath' rather than 'under oath'), independently of other frequency effects. For verbs, we computed both inflectional and relative entropy over inflectional paradigms. For nouns, we conceived paradigms as sets of prepositional phrases, and again computed the inflectional/relative entropy over 'prepositional paradigms'. The relative entropy effect is observed in single-word lexical decision data for Dutch, but not for words in narrative context. A discrimination learning model trained on the SoNaR-books subcorpus returned the same effect of relative entropy only at earlier training stages. This suggests that early benefits of highly discriminative contexts in predicting lexemic outcomes ('lexomes') are reduced as the idealized learner's corpus sample size increases. In other words, as the average number of outcomes supported by any one set of cues increases, there is a decreased likelihood that a certain paradigmatic context ('under', '-ed') cues a given noun or verb more than the content lexeme itself.