CLSP Fall Seminar Series: David Mortensen

Description

David Mortensen, an assistant research professor in the Language Technologies Institute at Carnegie Mellon University, will present a Center for Language and Speech Processing seminar titled "Extension Is All You Need (But Sometimes Good Intensions Help, Too)."

Abstract:

The philosophical distinction between intension and extension is relevant to many problems in mathematics and the sciences (including linguistics and computational linguistics). Intention refers to the definitional properties of a construct and extension refers to how a construct is manifest in the world.

This talk describes two computational linguistic studies which present seemingly opposed perspectives on the question of extension versus intention. The first concerns a family of sound patterns that are difficult to describe if one starts with articulatory features—phonological ordering effects in elaborate expressions and coordinate compounds. We show that these patterns, while frequently phonetically arbitrary, can be very robust and that they are learnable from surface distributions (by simple decision trees and SVMs as well as CNNs and LSTMs) without intentional supervision. We also show that there is an evolutionary pathway between an articulatorily coherent origin of these patterns and their current extension.

In the second study, we demonstrate that neural phonetic embedding methods supervised with universal articulatory features outperform methods without such supervision on many tasks in a comprehensive evaluation suite we have developed. On the other hand, purely data-driven approaches work best on tasks that are not defined in terms of articulatory features. These experiments suggest that intentional representations are useful just in case there is a clear mapping between intention and extension and adequate extensional representations are not available (as with the language-agnostic tasks in our test suite).