Re-thinking Labov's Pronunciation Drift Data

Continuing the discussion from Where Rick's Chapter 7 on "Social Control" goes off track where Bruce wrote:

On re-reading Labov’s paper I detected evidence that his model of the results is essentially equivalent to mine inasmuch as it assumes that similarities in pronunciation result from imitation. Here’s a quote from Labov’s paper that suggests that this is the case:

Only when social meaning is assigned to such variations will they be imitated and begin to play a role in the language. [emphasis mine-RM]

So Labov’s model seems to be that people tend to imitate pronunciations that have a preferred social meaning. His main evidence of this is the data in his Table 6:


A PCT model of this could be constructed but quite a bit more data is needed to make the model realistic. Actually, the first thing I would like to see is just a conventional regression analysis with variables he studied – age, geographical location, orientation towards Marth’s Vineyard, etc – as predictors and CI index as the dependent variable. This would give a more precise idea of which of these variables is actually the best predictor of CI index.

Labov’s conclusion that orientation toward Marths’s Vineyard is the best predictor of CI index (based on the data in Table 6) is flawed because it doesn’t take into account the correlation between that variable and other predictors; for example, those positive toward Marth’s Vineyard may live in one mainly in a particular region of Martha’s Vineyard and be of a particular age.

It would also have been nice if Labov had reported the medians and standard deviations of the CI measures as well as the means. Then we could tell whether an apparently large difference in average pronunciation, as is apparent in Table 6, is robust or a result of highly skewed data.

But the data that would be most relevant to an imitation model of pronunciation are measures of the degree to which the members of the population under study interact with each other.

So my model is a reasonable demonstration of the emergence of pronumciation differences that accounts for the available data. It’s defiitely a simple model. But I don’t think it’s correct to say that the model is too simple; the model isn’t too simple; it’s the data. But that’s certainly not Labov’s fault. He did great work but he was working from the perspective of the causal paradigm and he collected data from that perspective. And, of course, he explained it from that perspective as well.

What is a ‘social meaning’? What perceptions are controlled by ‘preferring’ one? Where in your model do preferences for social meanings appear?

I stand corrected. It does involve gradual averaging of Labov’s ‘centralization index’ abstraction (CI).

CI is a convenient abbreviation for a narrow range of variation in two perceptual variables, the center frequencies in the lowest two bands of amplified harmonics in the speech sound, called the the first formant and second formant. ‘Centralization index’ is not a common term of art in the field. Labov appears to have invented it for the sake of a compact representation in his tables.

These variables, F1 and F2, are co-varied over much wider ranges to produce and perceive the full range of vowel and consonant sounds. It is possible but very unlikely that speakers in this population developed a unique CI perceptual input function, and combining them into one higher-level ‘amount of centralization’ variable is not necessary to account for control of sounding like or unlike others. As a confound, greater or lesser centralization of all vowels is a function of gain (careful pronunciation vs. lax, unstressed pronunciation).

What is involved here is a change in the articulatory/acoustic reference values for the onset of the two diphthongs. Reference values may be achieved in careful, high-gain pronunciation, but typically are not completely attained, and control is typically disturbed by control of what comes immediately before and after (‘co-articulation effects’). Individuals displaying social identification with high gain control the socially differentiated reference value with high gain. The observed phenomenon of hypercorrection resulted.

But set that aside.

The model asserts that individuals approximate to one another’s pronunciation at each encounter. It is an obvious truism in linguistics that people often do this, but together with the contrary observation that individuals may differentiate features of their own pronunciation from values of those features as perceived in others, or they may exaggerate the observed values or extend them beyond the range in which they can be observed in others. Labov provides data on both of these phenomena, and both are important for the questions that he was investigating and for the explanation that he discovered. Those questions were, why did an extended period of gradual approximation come to an end and then reverse the direction of change, and why was the reversal generalized from the environment before a high back vowel (as in bout) to the environment before a high front vowel (as in bite). The investigation disclosed that the generalization took place in a subpopulation, adolescents, but only in that further subpopulation who intended not to move away. It is clear that this was due to their controlling higher level variables, and Labov clearly said so, although he did not use control theoretic terms.

If the individual’s pronunciation changes during or due to an encounter, it is because a higher level of control changes the reference values for the variables involved.

There is no such fragility in the data, including data from many other such investigations in New York and Philadelphia, and elsewhere by others. This defect in the model might be remedied by modeling higher levels of control adjusting reference values for formants, rather than making those changes an uncontrolled consequence of proximity or ‘encounter’.

The collective value for a group ‘suddenly diverging to a new value’ in real population speech data would be due higher-level systems correcting error in their input by changing references for these perceptions at lower levels. Such higher-level perceptions are amply represented in Labov’s paper.

Much of the language difference depended upon whaling terms which are now obsolete. It is not unnatural, then, to find phonetic differences becoming stronger and stronger as the group fights to maintain its identity.

What is it for a group to have an identity, and for members of it to control perception of that identity with high gain? (I hope you have no objection to that PCT translation of “the group fights to maintain its identity.”)

This gradual transition to dependency on, and outright ownership by the summer people has produced reactions varying from a fiercely defensive contempt for outsiders to enthusiastic plans for furthering the tourist economy. A study of the data shows that high centralization of /ai/ and /au/ is closely correlated with expressions of strong resistance to the incursions of the summer people.

Does this support a model of imitating those who you most frequently encounter?

That was Labov’s phrase so I don’t know know what he meant by it. The only point I was making by quoting Labov was to show that he, like me, believes imitation is the basis of the observed phonemic drift.

I think CI is a good hypothesis about a possible controlled variable, which is a function of the two other variables that happen to be more commonly used in the field.

I also think it’s unlikely that CI is a controlled variable but the actual controlled variable is probably very close to it. Evidence for how close CI is to being the correct definition of a controlled variable would have been the size of the variance of the CI values for each group; the smaller the variance, the more likely it is that CI is a CV, It would have been nice if Labov had reported the variance data.

That’s a nice point about gain (careful pronunciation vs. lax, unstressed pronunciation) being a possible confounding variable. One way to see if it was a confound would be to look at the variance of the CI values in the different groups. If gain, and not CI references, were what is different across groups then an indication of that would be low CI variance in the high gain groups and high CI variance in the low gain groups. And the average CI value in the high gain groups should be nearly the same since they are all keeping the CI value at the “true” reference.

Don’t these questions require longitudinal data – data showing variations in pronunciation over time? I don’t think Labov collected such data. Oh, and I don’t understand the second question. What does it mean for a reversal to be generalized from the environment before a high back vowel to the environment before a high front vowel? What is the evidence that a reversal has been generalized from or to an environment?

My model doesn’t rule out control of “higher level variables” as an explanation of Labov’s results. Indeed, the large pronunciation difference for people with a positive, neutral or negative orientation toward Martha’s Vineyard probably reflects control of a higher level variable that might be called “amount of socializing with people who share your attitude”.

Yes, that’s the way my model currently works, with the higher level controlled variable being “degree of imitation”.

This isn’t a defect of the model unless the other studies that you mention kept track of average pronunciation over long periods of time and found no sudden spontaneous changes. The sudden, spontaneous changes in pronunciation seen in the behavior of the model were a surprise and represent a prediction of the model. Now what is needed is a study of variation in pronunciation over many years. If no spontaneous shifts occur in that time – or if they occur for obvious outside reasons, such as a large change in who people interact with – it would be a count against the model.

I think a group identity exists when each member of a group is controlling for what all agree is essentially the same system concept perception. The people who do it with high gain are the “priests” or “leaders” of the group.

Maybe, maybe not. It is not necessarily inconsistent with it. If the people who express strong resistance to summer people tend to interact with people who feel the same way then my imitation model gets some support. If they don’t, then the model will probably have to be revised somewhat. But I think the model is on the right track. What is needed is the right kind of data to test it.

Labov’s data is a good start and all the data he reports can be accounted for by my imitation model. But to really test the model we need more data; variance measures, frequency of interaction between people in different groups, time variations in pronunciation, and so on. What we don’t need is “seat of the pants” theorizing based on what people deeply steeped in the conventional view of behavior have to say about what the data imply.