Computational linguists overall agree that morphology, with the exception of reduplication, is regular. Here regular is meant in the sense of formal language theory. For any given natural language, the set of well-formed surface forms is a regular string set, which means that it is recognized by a finite-state automaton, definable in monadic second-order logic, a projection of a strictly 2-local string set, has a right congruence relation of finite index, yada yada yada. There’s a million ways to characterize regularity, but the bottom line is that morphology defines string sets of fairly limited complexity. The mapping from underlying representations to surface forms is also very limited as everything (again modulo reduplication) can be handled by non-deterministic finite-state transducers. It’s a pretty nifty picture, though somewhat loose in my subregular eyes that immediately pick up on all the regular things you don’t find in morphology. Still, it’s a valuable result that provides a rough approximation of what morphology is capable of; a decent starting point for further inquiry. However, there is one empirical argument that is inevitably brought up whenever I talk about the regularity of morphology. It’s like an undead abomination that keeps rising from the grave, and today I’m here to hose it down with holy water.

Continue reading

As you might have gleaned from my previous post, I’m not too fond of features, but I haven’t really given you a reason for that. It is actually straight-forward: features lower complexity. By itself, that is actually a useful property. Trees lower the complexity of syntax, and nobody (or barely anybody) uses that as an argument that we should use strings. Distributing the workload between representations and operations/constraints over these representations is considered a good thing. Rightfully so, because factorization is generally a good idea.

But there is a crucial difference between trees and features. We actually have models of how trees are constructed from strings — you might have heard of them, they’re called parsers. And we have some ways of measuring the complexity of this process, e.g. asymptotic worst-case complexity. We lack a comparable theory for features. We’re using an enriched representation without paying attention to the computational cost of carrying out this enrichment. That’s no good, we’re just cheating ourselves in this case. Fortunately, listening to people talk about features for 48h at the workshop gave me an epiphany, and I’m here to share it with you.

Continue reading