The Grand Phonological Theory of Everything

[alice]

Is there a theory of phonology which is applicable to all known languages. i.e. it can elegantly explain all sound changes, allophonies, inventories, and so on? Straightforward feature theory (Jakobsonian?) can't cope with scalar-valued features like vowel height, and adding scalar values doesn't really work either, so I've been led to believe. Dependency phonology looks interesting, if a bit intricate, and there isn't much written about it. Are there any others?

[Frislander]

Word-level exemplar theory.

[Tropylium]

Straightforward feature theory (Jakobsonian?) can't cope with scalar-valued features like vowel height,

Note however that "scalar-valued" is itself an analysis here, not the base level of data.

Featural models of vowel height usually seem to go with /i/ [+high -low], /e/ [-high -low], /a/ [-high +low]. /ɛ/ can be, depending on the language, either [-high +low +front] (e.g. Vietnamese, Hungarian) or [-high -low -tense] (e.g. English; where the [-high +low +front] vowel is instead called /æ/). /ɪ/ seems to be pretty much always either [-tense] or [-ATR] (the IPA's concept of "near-close" strikes me as nonsense that has no systematic evidence behind it).

[Rory]

All known theories of phonology either overgenerate (i.e. they predict languages which are actually impossible), undergenerate (i.e. they fail to account for some kinds of attested phenomena), or both.

Stephen Anderson's 1985 book on the history of phonology characterized the field as moving between theories of rules and theories of representations. These considerations are largely what has driven advancement and changes in the field.

So, here's a list of some major theories of phonology. Several of them are not mutually exclusive, as they are separately theories of rules or theories of representations (e.g. there was recently an article published on doing strict CV phonology with optimality theory, just for kicks).

• Classic Chomsky and Halle's (1968) The Sound Pattern of English (SPE), a foundational text in generative phonology. Introduced many of the basic concepts that we take for granted today.
• Autosegmental / nonlinear phonology. This gives us a new type of representation, the idea that segmental information exists on a separate tier from autosegmental (usually tonal) information. This allows us to model tone interactions in African languages much more easily than with an SPE model.
• Feature geometry. The idea that features are organized hierarchically, rather than just being a list. This puts some constraints on the kinds of features we should expect to see interacting.
• Element theory. (This often goes with Dependency phonology.) Rather than features, which are always present and are binary valued, elements are privative (exist or don't exist) and combine in some complex ways to make sounds. I don't know much about this theory, to be honest.
• Optimality theory. Rather than applying a set of serial rules to derive an output, there are a set of ranked violable constraints, and the input-output mapping is simply a (parallel) optimization problem. This theory cleans up a lot of the weirdnesses of rules, makes strong links to linguistic typology (and phonetic naturalness), and fails to account for a lot of attested processes.
• Harmonic serialism. Originally proposed as a footnote in one of the first optimality theory texts, this theory has been revitalized recently. For the purposes of our discussion here, it's very similar to optimality theory.
• Gradient symbolic computation. A brand new idea, trying to link our understanding of neural networks (i.e. continuous mathematics) with our understanding of symbolic computation (i.e. discrete mathematics) and do phonology with it. I'm also not super familiar with this framework (yet!).
• Strict CV phonology. (Arguably this has its roots in government phonology, which grew out of dependency phonology.) This is a representational approach where sounds are arranged on a skeletal tier in alternating C and V slots. I... don't really understand this theory.
• Exemplar theory. Arguably not even a theory of phonology at all, this is an approach to representation which eschews symbolic underlying form in favor of detailed acoustic traces of words that you have heard. Rooted in theory from speech perception and cognitive psychology.

In my view, the real test for a theory's adequacy is how well it can deal with signed languages. This topic has not been covered much in the theoretical literature, to my knowledge.

[zompist]

Really interesting summary, thanks!

But why should a theory of phonology have to account for sign language? I mean, linguistics obviously should. But it's kind of like expecting acoustics and vision to work the same. There are similarities (both because of physics and because of neurons), but also important differences.

[alice]

Rory, that's very useful. It casts a lot of light on my original motivation for asking, which pertains to my ongoing quest to create a set of rules for my multi-language SCA which don't get stranded somewhere between individual but underspecified phonologies and global but overspecified phonetics. But I agree that, unless there's a one-to-one correspondence with phon[etics/onology], it's probably better to leave sign language out of it. Unless by "signed languages" you mean something different, such as "languages which can take negative values"...

[Rory]

But why should a theory of phonology have to account for sign language? I mean, linguistics obviously should. But it's kind of like expecting acoustics and vision to work the same. There are similarities (both because of physics and because of neurons), but also important differences.

Well, alice originally asked "Is there a theory of phonology which is applicable to all known languages" (emphasis mine). Sign languages are natural languages.

A theory of phonology should have to account for sign languages because phonology is not about sound. Phonetics is the study of speech sounds. Phonology is the study of the abstract (i.e. mental) units that words are composed of ("cenemes", if you want to roll with Hjelmslev's work). Now, it just so happens that these units have a strong connection to phonetics - and there's an ongoing debate as to "how much" phonetics is active in a given language's phonology - but phonology itself is not about sound. This might seem rather counterintuitive given the name of the subject, but I would wager that a great majority of phonologists would agree with me.

In the same vein, syntax is not about meaning. A well-defined syntax can exist for semantically meaningless systems. It happens that meaning is infused into every natural language syntax that we examine (and more and more, syntacticians are realizing the value of semantics in their theories), but the fields of syntax and semantics study different things.

Rory, that's very useful. It casts a lot of light on my original motivation for asking, which pertains to my ongoing quest to create a set of rules for my multi-language SCA which don't get stranded somewhere between individual but underspecified phonologies and global but overspecified phonetics. But I agree that, unless there's a one-to-one correspondence with phon[etics/onology], it's probably better to leave sign language out of it. Unless by "signed languages" you mean something different, such as "languages which can take negative values"...

Ah, if you're building an SCA, then there's a load of theoretical assumptions built in - words are composed of a finite set of discrete ordered symbols; the neogrammarian hypothesis is absolutely true (i.e. sound changes apply immediately to all words in a lexicon); sound change is regular and systematic; etc...

[Curlyjimsam]

Rory clearly knows more about this than me. Based in my work on syntax, though, I'm not sure a formal theory of phonology that overgenerates is necessarily a problem provided the unattested patterns can be ruled out by other, non-formal means - e.g. a particular system that is formally permissible might be unacquirable in practice due to the limitations of human perception and/or articulation. We also have to be careful about the distinction between "impossible" and "we just haven't actually observed it" (always a very tricky question).

[LinguistCat]

I'm not trained in this at all, but I have to agree with Curlyjimsam: A theory that under generates in some way is more or less proven false, but one that over generates might mean that we don't have enough data. I'd go for a theory that doesn't under generate but only over generates slightly or makes the fewest assumptions, but we also don't have a Theory of Everything in physics yet either, so I'd be happy with a "close enough" until we get enough proof to throw out competing theories or that theory is disproved.

Alternately, choose the one that isn't disproven but is the most interesting.

[Rory]

Based in my work on syntax, though, I'm not sure a formal theory of phonology that overgenerates is necessarily a problem provided the unattested patterns can be ruled out by other, non-formal means - e.g. a particular system that is formally permissible might be unacquirable in practice due to the limitations of human perception and/or articulation. We also have to be careful about the distinction between "impossible" and "we just haven't actually observed it" (always a very tricky question).

Yes, this is a good point. Historically, overgeneration was seen as more of a problem, as it was thought to be the "job" of formal phonology to explain typology - it was part of UG. Nowadays, a more common view is that a lot of typological trends are due to phonetic biases, and therefore aren't "part of phonology". Still, as alice implied in the original post, many people (rightly, in my view) consider that a phonological "theory of everything" ought to be able to account for how and why those phonetic factors influence phonology (as the effect is clearly not deterministic).

I'm not trained in this at all, but I have to agree with Curlyjimsam: A theory that under generates in some way is more or less proven false, but one that over generates might mean that we don't have enough data. I'd go for a theory that doesn't under generate but only over generates slightly or makes the fewest assumptions, but we also don't have a Theory of Everything in physics yet either, so I'd be happy with a "close enough" until we get enough proof to throw out competing theories or that theory is disproved.

Generally I agree with you, but often these overgenerating theories turn out to be unfalsifiable (or at least practically unfalsifiable), rendering them nonscientific. Sometimes we can learn more from a wrong theory than we can from one which isn't disproven.

[Salmoneus]

Well, 'unfalsifiable' is a trickier concept in "sciences" like linguistics, in which almost everything is unfalsifiable in a strict sense.
Ultimately, this is because linguistics is a narrative tradition - like botany, political science, history or theology - rather than an experimental process like the sciences in a narrow sense. Rather than making predictions tested by specific, independent experiments, linguistics attempts to fit the most congenial narrative to an (almost-entirely) already-existing field of data (even if it's not all been codified yet). Each narrative will - being a narrative and not reality - fail to be congenial in some way, but each narrative can be progressively elaborated or reframed (maintaining the same data) to appeal to critics. In other words, almost all of linguistics is the sort of post-hoc explanation that the falsifiability criterion was meant to exclude.

This is also why theories are much harder to kill off in linguistics, and why totally different theories can continue to survive in competition: there is no clear procedure for eliminating them, as judging the most congenial narrative is a subjective exercise (far more so than judging experimental failure) and there will always be wide divergences in aesthetic taste when it comes to narratives.

-------

Specific to this question: I think the remarkable thing is that anyone would ask it, and that nobody would question it. It seems to me that the entire attempt to provide a "theory" of phonology is misguided. Consider a closely-related subject: nobody would seriously ask for a theory of animals, which neatly explains all observed animals, neither overgenerating (predicting animals that don't exist) nor undergenerating (failing to predict the existence of the platypus).
Yet animals and languages are very similar things: both are complex structures that have emerged through evolutionary pressures acting upon stochastically-generated mutations over time, both are linked both through cladistic hierarchies of relatedness and through processes of convergent evolution, both are found in a bewildering (but non-random) diversity of forms arranged through space and time, and both develop almost entirely prior to the introduction of experimental observation and hence are conceptually "given". Indeed, languages are animals taken a step further, in the sense of being even less amenable to empirical study on account of greater degrees of indeterminacy and subjectivity.

Hoping to develop a "theory" that "generates" the features of a language from some universal is exactly parallel (only worse) to hoping to develop a "theory" that "generates" the platypus from some Ideal Animal. It's a bizarrely 19th century notion. Well, it's actually a bizarrely ancient greek notion (this sort of linguistics is really just an acceptable form of Platonism), but versions of it remained popular in some other subjects into the 19th century - think of the way Comte, Hegel, Spencer and Marx attempted to create "theories" of history that, through modulation of various parameters, could "generate" observed historical societies (societies are much like animals or languages in this respect - zoology, linguistics and sociology/history should probably be considered branches of a single discipline of, let's call it, "myriology"*). This approach has long since been recognised to be both infertile and confused in every area EXCEPT linguistics.

[Indeed, taking another branch of myriology: it's equivalent to asking for a "theory" of psychology that "generates" individual personalities... (neither over- nor under-generating!). Even psychologists no longer claim to do that, and they're not exactly paragons of rigour themselves.]


But what about the Universal Grammar (or whatever you call it when you don't want to call it Universal Grammar but want to continue as though it were a Universal Grammar)? Doesn't the existence of Universal Grammar entail the generatability of languages?

No, clearly it doesn't. Philosophically, universal grammar is only necessary in a very minimal form: there must be some general contours of more and less congenial assumptions about meaning (otherwise it would never be possible to narrow down the infinite range of possible interpretatons to the point where learning through affirmation and elimination becomes feasible), particular in the area of assent and dissent, but there is absolutely no need for, or even really plausibility for, a positive common grammar, or even for a determinate number of commensurable 'rules'. Beyond this is just faith, not reason.



*I'm calling it that because at core 'myriology' is the study of commonalities and differences between the myriad outcomes of common processes. It's not the best word, to be sure.

[alice]

That reads like a rather long-winded "no", Sal

Perhaps it's less confusing if I ask "has anyone come up with a phonological model which is applicable to all known (spoken!) languages, etcetera"? Or does that merely amount to the same thing?

[zompist]

Ultimately, this is because linguistics is a narrative tradition - like botany, political science, history or theology - rather than an experimental process like the sciences in a narrow sense.

Ah, the early 19th century... how is it there? I've always wanted to visit. The architecture seems so lovely.

It's the thrown-in "botany" which really elevates the statement from physics envy to full-bore antiquehood. Yep, that's all botanists do, they arrange specimens in glass cases and tell stories about them. No one has ever come up with theories of evolution, of genetics, of DNA. There is no such thing as philosophy of biology. Ernst Mayr never existed in this timeline. The only true science is physics.

Too bad even physicists have been lax lately in providing neat non-narrative theories that can be easily falsified. How long has string theory or many-worlds theory been increasingly elaborated without much evidence?

A little less ironically: I actually agree with you that a lot of science is classification and narrative. What I find foolish is that you wish to restrict "science" only to fields that minimize these things. The goal is not to make biology look exactly like the study of colliding billiard balls. The goal is to understand where biology differs from physics and understand what principles and methods are appropriate for it.

Hoping to develop a "theory" that "generates" the features of a language from some universal is exactly parallel (only worse) to hoping to develop a "theory" that "generates" the platypus from some Ideal Animal.

Yet that's exactly what biology has. We call it "genetics". There's this thing called a nucleus, see, and it has a little program inside which encodes the form of the platypus, or the nematode, or the tobacco plant, or whatever. The building blocks of DNA are sufficient to build all of these life forms; it's a marvelous and amazingly compact form of a generative theory.

You seem to think it's suspicious that you can't start with DNA and predict the platypus, but that's not what generative biology wants or claims to do. Amazingly, biologists start with the understanding that biological forms are enormously, bafflingly varied. The advance of biology has been in reining in the apparent pointlessness of this variation: finding reasons for it and finding limits for it.

Note that the largely correct narrative of Darwin was not backed up by the theory of genetics till nearly a century later. Generative linguistics is much younger, which makes it much more frustrating for outsiders... but rather exciting for people in the field. A wiseacre like yourself can come along and make fun of the idea of universal grammar... well, your 19th century friends also had a fine time making fun of the idea of evolution. UG may or may not exist, but it's interesting to look for it.

(To be clear, UG has been way oversold. Few if any universals really hold up... and indeed, some of the most seemingly safe and banal universals (like "all languages have vowels" or "linguistic signs are arbitrary") are, as Rory points out, falsified by Sign. But, well, "you have not solved every basic problem in your field" is not an interesting critique of any science whatsoever.)

[Frislander]

Also Sal have you never heard of Psycholinguistics? Or Sociolinguistics? Or Computational Linguistics? Or Language Acquisition Studies? Or Phonetics? Or indeed any field of linguistics outside of phonology, syntax and historical linguistics. If you look outside that little box of abstract theorising you'll find that applied linguistics actually does feature a large amount of experimentation, and indeed all linguistic theories are to some extent informed by experimental data. Sure the approach of applying the theory may appear to be different from the more traditional sciences, but really all it is is just applying the same basic principles like Occam's Razor to find the most parsimonious explanation for the phenomena in question, as with all science.