Well, I’m starting to feel in over my head with the flapping and specification discussion Eric and I are having. My interpretation of Sally Thomason’s informal study (reported here) led to this exchange, in part because Eric found some unintended implications in it, and in part because I had bandied about terms like ‘underspecification’ without clarity.

The term ‘underspecification’ carries several connotations: at its essence, it refers to a segment lacking specification for a particular feature (in its underlying representation). The utility of this sense of underspecification is in representational parsimony. Underpecification also has some theory-specific usage: in ‘Contrastive Underspecification’ (Clements 1987, Steriade 1987), segments are specified minimally with the features that distinguish them from each other. In ‘Radical Underspecification’ (Kiparsky 1982, Archangeli 1984, Pulleyblank 1986), segments are even less specified, down to the smallest set of features with which surface contrasts could be derived (via redundancy rules). If this distinction is too arcane or technical or unclear, let me sum up by saying that Radical Underspecification allows underspecified segments to persist within an extrinsic rule-ordered system, such that redundancy rules (which assign features to underspecified segments) can be ordered crucially after some other phonological operation.

Needless to say, dropping words like ‘underspecification’ can unleash a wide range of emotional reactions, especially when Radical Underspecification becomes involved, since this particular approach is difficult to reconcile with the non-serial OT (and also RotB). This may be why Eric doesn’t think it’s fun. (and when I said ‘fun’, I meant in a linguistic sense — the same way that hengma is fun. Twisted and unsuccessful phonology humour).

Anyway, what we arrived at is an agreed distinction between Lexicon Optimization and Lexicon Underoptimization, the latter of which I suggested as a way of removing redundant specifications from lexical representations wherever “traditional” Lexicon Optimization fails to do so. If Lexicon Underoptimization has any utility, it is in the removal of redundant specification from lexical representations — in the sense of theory-neutral underspecification, or at the very most, Contrastive Underspecification.

Underoptimization, I suggested, was responsible for preliterate English-speaking kids positing an underlying /t/ for invariant flaps. The implication of this claim was that /t/ is somehow “less” specified than /d/ and /ɾ/. This is (kind of) what I meant at the time, but Eric most recently points out that, empirically and substantively, it is difficult to support it:

Bob’s reasoning is that there’s something [d] has that [t] lacks: voicing. As well-accepted as this reasoning is, I find it hard to pin down its justification.

First of all, the idea that voicing is “something” and that voicelessness is a lack of that “something” is not as obvious as it may seem; both results involve laryngeal gestures (stiff vs. slack vocal folds) that could, for all we know, be “specified” on both segments. (This, by the way, is why I mistakenly latched on to the markedness thing before – it was the only possible justification I could see for assuming that voicing is something that voicelessness is not.)

Without having thought it through adequately, I let on that /t/ is less specified than /d/. If there’s any way for me to retract this implication, then I’d like to. But in thinking about it, it seems to me that Underoptimization or something like it might still be at work in the flapping issue.

In the original proposal, I claimed that when a learner encounters an invariant flap (as in medal), she has no evidence to choose between /t/ or /d/. So, evidently, her strategy is choose /t/. But the more elaborated version of Underoptimization would suggest that the learner chooses /T/: a coronal, anterior stop, unspecified for voicing. In natural speech, this unspecified stop becomes a flap, but in slow speech, it becomes [t].

Here’s where underspecification comes up … in pre-OT underspecification, there is a direct link between the features that are “missing” from an underlying representation and the redundancy rules that fill them in. For example, the redundancy rule [ ] –> [+hi] works in a system in which some vowels have no value for [hi]. (The decision you make as an analyst about which features are unspecified is tricky, hence Eric’s misgivings about contradictions and blind alleys and difficult assumptions about what phonological representations consist of).

But in this [t]~[ɾ] alternation, Underoptimization has given us /T/ as the underlying segment in ladder. We don’t know what it gives us for surface [t] and [d] though, hence Eric’s objection: how would Lexicon Underoptimization decide which specifications are removeable and which aren’t? Maybe I’m arguing myself into a hole, but it must be able to know what the output constraints are and how they’re ranked – just like the pre-OT Underspecification algorithm arrives at a combination of underspecified segments and redundancy rules. Instead of redundancy rules, we can now use constraints like this one, at least for the pre-literate speakers:

[+Obs] –> [-Voi]: Obstruents are voiceless.

Assuming Ident[voi] outranks this constraint, and in the absence (for now) of other phenomena, the result is that /t/ and /T/ converge on [t] as an output, while /d/ comes out as [d]. (And as an alternative blind alley, we could instead appeal to [+Obs]->[+voi], in which case /T/ converges with /d/).

So, Lexicon Underoptimization needs to ‘know’ that [+obs]->[-voi] is ranked below Ident[voi] in English, in which case, it is OK to posit /T/ as an underlying segment. How it can ‘know’ this is a modelling issue which might ultimately sink the proposal, or might fuel future posts.

Meanwhile, other phenomena come into play, for example, the flapping contraint FLAP which renders /t/ and /d/ to [ɾ] foot-medially before unstressed vowels. FLAP outranks Ident[Voi], which is how we get flaps in words like writer. It thus also outranks [+obs]->[-voi], which is how we get flaps in invariant words like ladder.

But then in the slow-speech task, something happens … I say, ‘slow speech’ makes every vowel stressed, removing the environment in which FLAP can apply. The result is that underspecified /T/ comes out as [t] in slow speech for any invariant flap, like in ladder, for anyone in which [+obs]->[-voi] is active. The really interesting test would be to ask the same little guys to pronounce rider real slow.