Lecture: Grammatical Factors in Morphological Processing
Evidence from Allomorphy
This study examines whether sensitivity to allomorphy impacts processing. English has two kinds of allomorphy: rule-based (cat/s/~dog/z/~hors/əz/) and suppletive (good~better~best). We investigated allomorphy as a novel predictor based on this distinction: first paradigm-by-paradigm (Experiment 1) and then across paradigms, stem-by-stem and form-by-form (Experiment 2). Mixed-effects regression models were fit to the BLP dataset (Keuleers et al. 2012) with standard predictors and RTs and accuracy as targets. Paradigms and stems both yielded the pattern (lowest to highest) none>rule>suppl for RTs, while inflected forms exhibited rule>none>suppl. Frequency interactions reversed the patterns, yielding suppl>rule>none for paradigms and stems and suppl>none>rule for inflected forms (all p<0.001). While accuracy patterned the same, the models failed to converge in both experiments. The findings show that sensitivity to allomorphy has to be addressed in processing models. However, further investigation of allomorphic variables is needed.
Trials from the British Lexicon Project lexical decision dataset (Keuleers et al. 2012) were annotated for allomorphy type along the cline suppletion>rule>none, with every paradigm member assigned only its paradigm’s highest complexity value for Experiment 1 (see Table 1 under the supplementary tables link). Mixed-effects regression models were fit to the by-trial dataset with standard predictors. Z-transformed RTs and accuracy ratings were separately modeled as the dependent variables. Frequency, paradigm allomorphy and their interaction were all found to be significant predictors (all p<0.001). Members of non-allomorphic paradigms elicited shorter RTs than allomorphic paradigms; within the allomorphic paradigms, members of suppletive groups elicited longer RTs, yielding the pattern none>rule>suppletion. However, interactions with frequency reversed the pattern, yielding suppletion>rule>none (Table 2). While the same general pattern was observed for accuracy, the models failed to converge.
The paradigms were then split into two additional predictors, one for stems (HasAllos) and the other for the (possible) allomorphs themselves (IsAllo). These were introduced into the models for Experiment 2 instead of the paradigm predictor. HasAllos yielded none>rule>suppletion, while IsAllo exhibited rule>none>suppletion (all p<0.001). Again, the frequency interactions reversed the pattern for both allomorphy predictors (Table 3). The accuracy models exhibited the same general pattern, but again failed to converge.
These findings indicate that not all information relevant to processing is distributional (cf. Baayen et al. 2011 and Marantz 2013), and sensitivity to allomorphy has to be addressed in models of lexical processing. Additionally, we have observed that findings for whole paradigms can be decomposed into individual forms. Possible routes for further investigation include crosslinguistic replication (replication in Dutch underway) and establishing an explicit lookup model in the mental lexicon.