CLASP
The Centre for Linguistic Theory and Studies in Probability

Dendrophilia squared

Abstract

Fitch (2014) suggests that at the heart of syntactic cognition is ‘dendrophilia’: “a propensity by our species to infer tree structures from sequential data”. This is an appealing idea but raises two issues. First, how precisely does this happen, and which tree structures should they infer, given the indeterminacy of the tree structures given the observed sequences? Secondly, trees are notoriously inadequate for syntax: a majority of syntactic theory is concerned with deviations from hierarchical structure — described variously as movement, displacement, internal merge and so on.

In this talk I will try to answer these questions using two ideas. First I use Roger’s notion of multidimensional trees (Rogers, 2003) as syntactic structures, generalising the relation between a string considered as a 1-d structure and a tree considered as a 2-d structure to higher dimensions. Secondly I take a model for learning tree structures from sequential data with PCFGs, (Clark and Fijalkow, 2020) and show that it can be extended to learning 3d-trees from 2d-trees (Clark, 2021) and above if needed. By applying this model of dendrophilia twice we can then, in principle, learn these highly structured 3d trees from strings.

I then relate this to discussions of “well-nestedness” in grammar formalisms, and the extent to which this can account for observed syntactic structures in natural language corpora.