sound change rules

[Daniel]

Hmm. What would it mean to "understand" a coin flip? I suppose in that sense what you're describing is possible.

On the other hand, how advanced is this research? Does it resemble something that actually works? Again, if you apply it to Middle English does it fairly often predict something like the Great Vowel Change?

I'm also uncertain about how much is conditioned in that way and how much is truly random based on social factors (and arbitrary variation that happens to be associated with them).

[MalFet]

Hmm. What would it mean to "understand" a coin flip? I suppose in that sense what you're describing is possible.

Do you understand what Markov models represent? If so, why are you comparing them to a stateless probability event?

[jkpate]

Ah, I've been occupied with travel but want to chime in real quick.

It's a Markov model in every sense, but to call it "meaningless and uninformative" strikes me as pretty bizarre.

It depends on what we mean by "meaning" or "informativeness". I'm talking about it from the perspective of the researcher understanding from it. ... If it turns out that linguistics (or parts thereof) are just probabilistic, then I'm not too interested in studying them

I remember sharing this sentiment when I was first beginning in Linguistics. I knew a bit of stats, but felt that probabilistic approaches were at best an ugly kludge to just get the right answer rather than a route to deeper understanding. This sentiment changed when I realized that probability distributions are richly-structured objects in their own right. What did it for me was the work described in Johnson (1997), which showed that simply conditioning on more or less information could dramatically change probability distributions. That is, there is a direct connection between how linguistic structures are put together and how probability distributions over those structures behave. Saying that something is "just probabilistic" is glossing over an enormous range of interesting and structured behavior.

Another resource you might consider is Sutton and McCallum's tutorial introduction to Conditional Random Fields, with particular attention to the difference between naive Bayes and logistic regression, and what happens when you plug these classifiers into each other (so that the predicted class of one classifier is the "observation" of another classifier). Alternatively, I haven't had a chance to watch the talks myself yet, but I just saw that Christopher Bishop has three Machine Learning Summer School lectures on graphical models available via youtube. (The other videos are by big names too, and look to be worth watching as well, but probably won't speak as directly to what it means to be a structured probability distribution).

Getting a handle on this all means sitting down and understanding the math, but I promise it is worth it if you want to understand what the models MalFet is talking about mean.

[Daniel]

I don't have anything against statistics. I just don't see it as the same kind of understanding as other approaches. Maybe it all ends up just being about whether it's deterministic or not, although when phrased like that I don't explicitly agree with that sentiment either.

Having an algorithm that can run a simulation of language isn't too bad-- it does explain something in a sense, or at least it imitates it relevantly.
But it also seems to miss the social factors-- who in the social network is prominent and what distribution of the variation do they happen to contribute? Historical change is, I think, more than just optimization of the system. Sometimes that may "explain" it (post hoc), but it isn't the full story.

[MalFet]

Who said anything about optimization? Markov models are fundamentally atelic. That's the whole point.

Also, isn't it a bit disingenuous for you to be making claims about what stochastic models do or don't "miss" without having actually read any of them? The book I recommended to you just a few posts up, for example, could be summarized as a diachronic application of Labovian sociolinguistics. If you think it's ignoring social dynamics, you're not understanding the argument.

[Daniel]

What's the tangible result from this approach? Pick a situation (real or hypothetical), and what would you say is the answer to "why"?

[MalFet]

Palatalization of velar obstruents in Salish: coarticulation with front vowels causes advancement of the tongue body, which favors articulatory patterns that are perceptually difficult to distinguish from palato-alveolar affricates (see Guion 1998 for experimental data). New learners acquire this as a form of allophonic variation (potentially free, potentially socially conditioned), which has the cross-generational effect of first expanding the scope of acceptable phonetic realization before resettling on exclusively palatalized forms.

There's an obvious typological lesson here, too, explaining why the palatalization of voiceless obstruents is vastly more common cross-linguistically than the palatalization of voiced obstruents, but that's getting a bit beyond the scope of the issue here.

[Daniel]

I like that answer. It doesn't seem entirely categorically different from answers like "t became d due to lenition", just much more sophisticated and better defined.

How much of that is post hoc-- explaining events after seeing them, and how much is actually predictive, given a more general case "why" for the current state toward a next state? I know those are similar, but to me what you said is a perfectly reasonable case of figuring out what happened, but it doesn't sound so different from "if this happened, then it was probably due to this". Is that a fair characterization, or does this line of research get beyond that?

Looking at a change and figuring out what it probably came from is one side of the puzzle. The other is looking at a language and figuring out how it will probably change. If both questions can be answered (at an above chance level), that's a first step toward the "why" question.

It also strikes me, in the typological generalization perspective, as something like "grammaticalization" theory, where some paths are more common than other paths (although in this case it's based on phonetic/articulatory information rather than cognitive/conceptual factors). The trouble is that it is difficult to distinguish frequent sequences of changes from actual cause and effect, although the implications are very interesting.

[MalFet]

You're still thinking about this in terms of predictive versus post-hoc explanations, but that's a fundamental misunderstanding of how non-deterministic phenomena work and how Markov systems represent them. There is no post-hoc here. Post-hoc is not even a conceptual possibility in this framework. There can't be a post-hoc because these are non-telic accounts of emergent phenomena.

The material mechanisms of change are, generally speaking, validated by both typological and experimental data. But, ultimately, since linguistic change is an aggregate phenomenon of populations, the question of cause is dramatically overdetermined. What an explanation accounts for are pressures on the system, and that's true of virtually all non-trivial accounts both inside linguistics and in other fields.

In other words, the idea that a good account should take Middle English and predict the Great Vowel Shift "fairly often" is completely misconstruing the nature of language. It's not something that happens fairly often. That's what makes it interesting. A model that predicts it a likely transformation out from Middle English would just be incorrect.

[jkpate]

Suppose we wanted to model Texas Hold 'em poker. There are many sources of randomness, including the shuffled deck, individual player propensities for bluffing and risk-taking, and individual player assessments of other players. So there is a lot of room for explaining the factors that contribute to various player decisions and hand outcomes. However, each individual outcome is low-probability, and the prospect of predicting the outcome of any particular deal is hopeless.

I think djr33's concern can be addressed with some discussion of overfitting. When fitting a statistical model, it's possible for the parameters to capture some regularities that happen to exist in the training data but do not generalize to new observations. Insisting that your model predict the actual outcome with high probability, however, encourages overfitting: you want the model to put an appropriate amount of probability mass on plausible events that, by chance, did not occur in the training data. An appropriate metric for this would be something like the perplexity of a held-out test set, which measures how well the predicted probabilities of events match their true probabilities.

[Daniel]

Post-hoc is not even a conceptual possibility in this framework.

Post-hoc is a kind of analysis. Surely any kind of analysis can be applied to a certain data set after the results are known (or before). My point is that there's a distinction between an analysis that lines up well with what we know (consistent with data post-hoc), and an analysis that expands our knowledge on its own.

There can't be a post-hoc because these are non-telic accounts of emergent phenomena.

Then perhaps the "why" question simply doesn't apply at all-- if the phenomena don't really exist anyway, then the "why" is just because it's an illusion, and the real question is what is underlying in the system. That may be a fair point.

In other words, the idea that a good account should take Middle English and predict the Great Vowel Shift "fairly often" is completely misconstruing the nature of language. It's not something that happens fairly often. That's what makes it interesting. A model that predicts it a likely transformation out from Middle English would just be incorrect.

Ok, now we're getting somewhere. I can certainly see that argument. But then how/why would anyone attempt to answer the "why" question? It seems like the answer is always some form of "an unlikely change occurred compared to the state of the language". This goes back to what I said: we can't really answer a "why" question at the level of human understanding/knowledge. There may very well be an algorithm that can describe what happened, but to answer a question like "why does [a] become [e]" seems to still not have a good answer. In fact, you just said it-- the best we could say would be something like "it probably wouldn't". And we can add some correlated factors (like proximity in the vowel space) to say "why", but it's far from something concrete (perhaps a deterministic explanation).

On the other hand, I'm not sure I agree: do you think that what happened in Middle English was due to chance, if you consider all variables? I know that's a philosophical question, but to me it seems like some properties of ME and its speakers actually determined that result, and it would probably repeat itself given the exact same situation. So the "why" question then could in theory be answered but by omniscience in a relatively uninteresting way.

For now I think I'll stick with what a colleague says-- "Shift happens!" (And there are various ways we can model it and understanding contributing/correlated factors like articulatory variability.)


jkpate, that sounds like exactly what I'm thinking here (though I wasn't thinking of it in those terms). Thanks. And what does that say about the "why" question?
To me, any "why" question regarding poker would be post-hoc-- why was it good to get the two of clubs? because in the end it worked. Or why did I win? because I got the two of clubs. But trying to understand the system beyond that would involve, basically, counting cards (and knowing the mental states of the players). So I'm not convinced it's very useful to ask a "why" question for poker or language change. Does that seem reasonable or unreasonable to you?

[jkpate]

jkpate, that sounds like exactly what I'm thinking here (though I wasn't thinking of it in those terms). Thanks. And what does that say about the "why" question?
To me, any "why" question regarding poker would be post-hoc-- why was it good to get the two of clubs? because in the end it worked. Or why did I win? because I got the two of clubs. But trying to understand the system beyond that would involve, basically, counting cards (and knowing the mental states of the players). So I'm not convinced it's very useful to ask a "why" question for poker or language change. Does that seem reasonable or unreasonable to you?

It does not sound reasonable to me. The whole point of a model is to draw more general conclusions than "because in the end it worked." It could work for very different reasons, depending on what was going in a particular hand, and what had been observed to work in the dataset as a whole. For example, in a particular deal, perhaps the flop had a 2 of spades, and the winner rarely bluffs, so the winner was able to pull a convincing bluff when a 2 of clubs was drawn and there were two face-up 2s. To draw this conclusion, we need to aggregate different statistics across multiple observations. A model specifies how we aggregate statistics. In this example, there are conclusions that are general across players ("a bluff is x% more likely to be successful if the bluffer rarely bluffs" and "a bluff is x% more likely to be successful if <certain card combinations> are face up"), conclusions that are specific to a player across hands ("Steve rarely bluffs", "Steve is x% more likely to bluff if it involves going all in", "Steve doesn't care who has the big blind when it comes to bluffing"). In other words, a model allows us to determine whether, and the extent to which, variables relate to each other across observations. Probability distributions are not just a mud of unstructured correlations.

For language change, this could take the form of identifying that some kinds of changes are related to  communicative variables, other (potentially overlapping) kinds of language change are related to maintaining or breaking down social strata, geographic adjacency to other languages with particular features, and so on.

[Daniel]

I still see that as answering a how question. (And that is important, scientific, and interesting.)
But I think often why is confused with how.

I think this either means:
1. We can know the HOW, but the why is always because "shift happens".
Or:
2. In the case of language change, HOW=WHY.

Perhaps you're of the second perspective?


In fact, this might be a more general situation in science: Why? Because. How? [insert long scientific explanation here]

There's wikihow.com but no wikiwhy.com... seems fitting?


"How" can be answered with a story. Science (in the best case scenario) gives us accurate stories about how the world works.
"Why" can be answered by motivating that story in a deeper way-- you'd need to ask a literature studies researcher "why" literature works, which ends up, I think, circular, possibly with "because we believe it"-- so the why of language change is the why of science-- because scientists believe science.

Or, we can just conflate how and why, and that might be correct.

But I've always found this question, when someone asks me, very strange: "Why did English develop like it did?" (Or any similar question.) I can certainly give a story about how English came to be, with its strange spelling and so forth, but I feel that I have to stop there, even if I'm just speculating-- I can't imagine a "why" answer.

[freknu]

I think you're making a far too narrow (and probably colloquial) interpretation of "why".

• A, B, and C speak α.
• A and B decide to shift all /a/ to /e/ to create β.
• A and B still speak α, but C does not speak β.

Why did /a/ shift to /e/?

Now, you might say that the decision by A and B to shift all /a/ to /e/ is the "why", but that actually says nothing about why /a/ shifts to /e/. If you instead focus on vowel space, vowel harmony, fronting, palatalisation, etc., then that will tell you "why" — even though to colloquial interpretaton it may not appear to do so, as it doesn't have an anthropomorphic agent and/or cause which is so often confused with "why".

The decision by A and B are environmental pressures, only due to their anthropomorphism they are given more weight and "reason" compared to "random" environmental pressures.

So, why did /a/ shift to /e/?

[Daniel]

I think you're making a far too narrow (and probably colloquial) interpretation of "why".

I'd actually think it's the opposite-- that colloquially we use "why" very broadly. Applying stricter conditions on it is my scientific agenda. If I casually ask you "why did it change" I might be happy with a response like "because the vowels are close and speakers started pronouncing them differently". If that's all that is required for a "why" then it doesn't seem like a hard question at all.


So do you see a distinction between how and why?