In a discussion of the pre-OT issue of Radical vs Contrastive Underspecification, Kenstowicz (1994:508) provides what turns out to be an understatement for this week’s flapping discussion:

Finding solid evidence bearing on this issue has proved to be very difficult.

In other words, it is difficult to find empirical support for positing underlying representations that lack any specification, plus or minus, for a particular feature. And indeed, the slow-speech test that led to this discussion doesn’t clearly identify the underlying representation of the consonant that the flapping rule normally obscures. So when Eric writes

The ranking of markedness constraints handles the compelementary distribution with no need for underspecification. There may be other reasons to adopt underspecification, but this is just not one of them.

I agree. I also want to reaffirm in public my subscription to the idea that surface generalizations are better analyzed with output constraints and not input constraints. But while we’re talking about underlying representations and a (perhaps overly) brief proposal I made in a previous post, I thought I’d try to clarify my point.

Underoptimization, as we’ve been calling it, is what I proposed as an alternative to Lexicon Optimization, as a means of determining an underlying representation wherever several converge upon the same output representation. The brief proposal is repeated below:

Maybe there is another way of addressing the lexicon that is more compatible with Underspecification – perhaps like choosing the candidate underlying form that has the fewest featural specifications.

Eric sees two problems here: incompatibility with RotB, and a failure to predict an underlying representation that differs from the one predicted by Lexicon Optimization:

A formal implementation of Bob’s Lexicon Underoptimization alternative, on the other hand, would (also) have to allow markedness constraints to evaluate inputs in order for /t/ to be considered less marked than /d/ as an input. But note that this procedure would still select /ɾ/ as better than both /t/ and /d/: in English, [ɾ] is less marked than both [t] and [d] in the context of flapping, which is why we have flapping in the first place! (emphasis added – bk).

The highlighted statement mischaracterizes Underoptimization. Something evaluates inputs, but Markedness is not it. I hadn’t intended markedness (formal or otherwise) to come into the equation, and I didn’t mean to imply that underlying /t/ or /T/ is less marked then /d/ or /ɾ/. Eric adds:

It’s this intuitive sense of “markedness” that leads to the widely-accepted idea – echoed in Bob’s post – that degree of underspecification is a good way to represent relative markedness. A [t] lacks a feature that a [d] has, and presumably both lack a feature (or features) that [ɾ] has.

That idea may be widely accepted, but I deny echoing it. The consonants in question may differ in their relative markedness in output representations, which is contextual: some are marked word-initially, and some are marked before unstressed vowels. However, a markedness constraint as I understand it cannot evaluate /t/, /d/, or /ɾ/ as underlying representations, since it evaluates output representations. It’s this property of markedness constraints that allows Lexicon Optimization to work as it does.

To pursue the example, let’s agree on the operative constraint, FLAP:

FLAP: [t] and [d] cannot occur foot-internally before unstressed vowels.

This constraint is satisfied by the output of ladder, [læɾɚ], regardless of whether the underlying form has a /t/, /d/, or /ɾ/, because the output has a flap before the unstressed [ɚ]. Lexicon Optimization predicts that the most faithful Input-Output mapping is the optimal one, so we end up with /læɾɚ/ as the underlying form. This is illustrated in the following ‘reverse tableau’, where different URs of the same output are compared.

To reiterate, each pairing satisfies FLAP, since they all have an output form that satisfies it. This is true whether you subscribe to Lexicon Optimization or not. But if we want to pick either of the other non-flapped underlying forms as the proper one, Lexicon Optimization does not provide a means of doing so. This was the point of the Underoptimization proposal: something must be able to compare candidate underlying representations, since output Markedness constraints cannot, and since Faithfulness constraints make poor choices.

The principle at work in Underoptimization is one that tries to remove feature specifications wherever possible:

Underoptimize: when positing underlying representations, remove any feature specification that is not needed to generate the proper output.

This is the same principle you’d use if you wanted to keep aspiration from being specified on foot-initial voiceless stops. Crucially, it is not the same as choosing the ‘least marked’ underlying form, and it’s not the same as using a Markedness constraint to evaluate an underlying representation. It also chooses an underlying representation of ladder in which the intervocalic consonant has no voice specification.

Ultimately it matters little, since we agree that ranking of markedness constraints is appropriate to handle surface phonological distributions. I just didn’t want the Underoptimization proposal to suffer from any mischaracterization derived ultimately from my earlier failure to elucidate it adequately.