Over at headsup: the blog, fev recently pondered variation in transcription practice ("Annals of g-droppin'", 6/6/2011). He starts by noting that the same paper edited the same quote, in the same AP story, to have -in' in some but not all gerund-participles in one version, but -ing throughout in another version. And his main concern is with the socio-political subtext of the choice to use eye-dialect in some cases and not in others:
It's worth puttin' the question to the AP (and/or your own political writers). What exactly are you trying to show about Palin's speech, and how consistent can you credibly claim to be about it — either within a single sentence of hers or among candidates who may have those or other speech features despite their necktie-wearin' formality?
We've had quite a few posts on this dimension of linguistic variation and its socio-cultural associations. A sample:
"The internet pilgrim's guide to g-dropping", 5/10/2004
"Empathetic -in'", 10/18/2008
"Palin's tactical g-lessness", 10/18/2008
"Pickin' up on those features also", 22/29/2008
"Pawlenty's linguistic 'southern strategy'?", 3/17/2011
"Symbols and signals in g-dropping", 3/23/2011
I need to confess, though, that the evidence in some of these posts is barely a step up from anecdotal. Thus in my post "Empathetic -in'", I wrote:
Other things equal, the rate of "g-dropping" goes up with lower social class; goes up with greater informality; and goes up with more positive affect (e.g. joking vs. arguing). Without using a loaded word like "slumming", and indeed without raising the question of consciousness at all, we should note that every one of us is making many choices about self-presentation every time we open our mouth, and in particular we add a brush-stroke to our self-portrait every time we choose a pronounciation for the English gerund-participle suffix -ing.
In the first 40 minutes of the first presidential debate, Senator Obama used 84 gerund-participles, and dropped 8 g's. A g-dropping rate of about 10% is not at all out of line for someone in his position — in comparison, in the same period of the same debate, Senator McCain dropped 10 g's in 66 opportunities. (In both cases, I've left out all instances of the sequence "going to", which is especially interesting but also behaves in a special way.)
N of 84 or 66 is not too bad — but my next sentence was
The key thing, though, is not the percentage but the positioning.
And as we start to divide up the instances according to various aspects of the context, the numbers get small in a hurry.
So I'm happy to say that Jiahong Yuan and I have demonstrated, at least in principle, the feasibility of automating the classification of gerund-participle pronunciations. Our poster from a recent speech production workshop is here: "Automatic Detection of 'g-dropping' in American English Using Forced Alignment".
We started with 200 +ing words randomly selected from the Buckeye Corpus, 100 that were phonetically labelled with [ ih n ] and 100 that were labelled with [ ih ng ]. Since those labels were imposed by a human listener, and human phonetic judgments are by no means intersubjectively uniform, we had 8 other native English speakers (with at least some phonetic training) perform a forced-choice binary judgment on each of the tokens.
We then asked a speech-recognition system to make the same binary forced-choice judgment on the same tokens. Overall, the human listeners (including the Buckeye transcript) agreed with one another, on average, about 86% of the time; the automatic procedure agreed with the human listeners, on average, about 85% of the time. (A detailed matrix of pair-wise agreement is here.)
We can no doubt improve this by doing some categorization-specific discriminative training — both for the humans and for the machine — but it's already good enough to go forward with. Specifically, we'll be able to automatically classify g-dropping behavior in arbitrarily large collections of transcribed political speeches and debates.