This week’s colloquium talk operates at the intersection of linguistics and machine learning. Colin Wilson of John Hopkins University will present his work on learning morphological mappings with neural networks.
The talk will take place Friday, September 21, at 3:30pm in SAC 305.
Title: Learning morphological mappings with interpretable neural networks
Abstract: Deep neural networks have proven to be highly successful at learning many linguistic patterns, including morphological mappings (e.g., work → worked), but their internal representations and processes remain largely opaque. In this talk, I describe a class of encoder-decoder networks that contain explicit representations of linguistic structures, such as stems and affixes, and whose processing steps can be straightforwardly understood as gradient application of elementary symbolic operations (e.g., reading and writing segments, overwriting of vowel melodies, stem extraction). Because all of the computations of these models are differentiable, they can be trained by gradient descent or other numerical optimization algorithms. Across a wide range of case studies — including operations of infixation, transfixation, and reduplication that have been challenging and unwieldy for previous models — the networks succeed in learning accurate and linguistically-interpretable analyses from unannotated examples of morphological mappings (e.g., [work, worked], [tristi, trumisti], [ʃamar, ʃimri], or [bagawen, babagawen]).