Language learning as expectation: a discriminative perspective

How do children acquire a productive language system? Traditional approaches assume this involves acquiring symbolic rules operating over discrete word categories, which has been argued to be “unlearnable” unless the principles of language are innate. In contrast, Liz’s work suggests that linguistic productivity is a direct function of language input (e.g., Wonnacott, E. et al. (2012), Wonnacott, E. (2011), Wonnacott, E., Brown, H., & Nation, K. (2017), Samara, A., Smith, K., Brown, H., & Wonnacott, E. (2017).

These findings can be described in terms of a balance between “item-based” learning – where structures remain associated with particular words – and generalisation. This interpretation nicely fits with Perfors, A., Tenenbaum, J.B., & Wonnacott, E. (2010)’s computational work, and also other input-based approaches – e.g., Tomasello et al., (2000). However, the explanatory value is limited in that there is no clear account of the underlying mechanisms of learning.

This research project stems from this shortcoming by outlining a perspective whereby learning results from environmental cues reducing uncertainty about outcomes. For language, the “outcomes” are a system of linguistic form contrasts, “cues” come from the world and earlier parts of an utterance, and “learning” is dissociating the (huge) set of uninformative cues and mastering the system.

Our approach is learning-theory inspired: we use the Naive discriminative learning (NDL) model in order to make predictions about how outcomes are predicted based on the cues available. NDL in combination with Artificial Language Learning experiments –exposing participants to experimenter-designed miniature languages – allow us to test key predictions about the effects of the input’s distribution and linear order on generalisation.

This is a four-year project in collaboration with Michael Ramscar (personal web-page) and is funded by a 2019 Leverhulme Trust Research Project Grant awarded to Liz. Eva Viviani is currently working as postdoc on this project.

More details to come.