What corpus linguistics can learn from neuroscience

Simon Dennis
University of New Castle, Australia

Two recent results from neuroscience can provide useful constraint for models of linguistic processes. Firstly, it appears that time is represented (at least in some areas) as a population code (MacDonald, Lepage, Eden & Eichenbaum, 2011). Secondly, it seems that as one moves from auditory cortex to the prefrontal cortex linguistic structure is captured by a set of temporal windows of increasing duration – analogous to visual receptive fields (Lerner, Honey, Silbert & Hasson, 2011). In this talk, I will describe a large scale model of text processing inspired by these observations. The model proposes that we can think of linguistic processing as a large structured recurrent network, with weights defined by pointwise mutual information. The model will be demonstrated on a question answering task.