Modeling lexical biases in morphosyntactic alternations: aligning usage-based theory with multilevel/hierarchical models
Abstract When language users choose between alternative schematic constructions, as for example in the English dative alternation, they are influenced not only by contextual or processing-related factors that apply generally but also by individual lexemes that fill the open slots of constructions. The methodology for modeling such lexical biases is still under development. To motivate methodological decisions, we need a theory of how lexical biases arise in the first place. Drawing on usage-based theory, this paper argues that lexical biases emerge from exemplars of concrete, lexically specific instances, from which generalizations are formed. The resulting hierarchical structure enables speakers to learn lexical biases efficiently by drawing on expectations derived from these generalizations. This assumption parallels the logic of multilevel (hierarchical) modeling, in which the effects of random groups are estimated jointly under a shared distribution, while also being separately estimated for each group. In order to support this claim, the present paper revisits the English dative alternation. We employ a Bayesian multilevel regression model in order to demonstrate its viability for capturing the lexical specificities of alternating constructions. Our case study serves to illustrate technical aspects of the method, such as specifying the group-level effects structure and setting priors.