What contemporary theories attempt to explain why languages have phonotactic restrictions instead of permitting any phonemic combinations?

Question

https://en.m.wikipedia.org/wiki/Phonotactics

As to why we see some combinations of phonemes allowed vs. not allowed, in some human languages, like how in English we have the phoneme clusters “sn”, “sm”, “str”, but not “sd-“ or “sq-“ or “[abc]” -

1. What current theories offer explanations for why language does not permit arbitrary combinations of all available phonemes, ie, all possible combinations, so in English, every combination of 25 letters of length n would be a known word with a meaning, like “ac” (snake), “aaa” (bird), “bgqp” (waterslide), “hjpmnxyo” (cousins), etc.?
2. Are there phonotactic theories that extend to non-adjacent (“distal” or “dislocated”) phonemic distributions, for example, “sn-“ calls to mind to me, (while writing this), words with some unvoiced consonants and some plosive consonants, “snake”, “snipe”, “snack”, “sneer”, “snide”, “snow”, “snot”, “snuff”, “snafu”, but not further consonants like “s” and “n” (alveolar nasal), like “snasna”, “snan”, “snunny”, etc.

Physically, the latter words strike me as exertive, like they do not feel natural in the mouth, but I have not established a correlation, and correlation ≠ causation.

In light of these questions:

There appears to be a general paradigm I am exploring in linguistics in general, which is sort of about “free generation” vs. a heavily pre-constrained rule system a priori to dictate what forms are generated (I believe the latter is much more similar to Chomsky’s idea of a “universal grammar”, but I do not know). Examples of the former might be thinks like link grammar and https://en.m.wikipedia.org/wiki/Construction_grammar, discriminative learning (here), and I think Chomsky’s Minimalist Program.

The question is sort of about how “particular” the “meta-“grammar (the constraints on whatever system or algorithm generates the grammatical forms of languages we end up observing empirically, like English and Spanish) of human languages may be. Some paradigms in linguistics - I think like Chomsky’s Principles & Parameters approach (1) - are heavily axiomatic, like they insist on a lot of structure, concepts, and rules as the core features of language / the language faculty (“LF”), like “each argument is assigned one of the follow theta-roles”. It seems to be implied that if language has such particular structure, somewhat similar to human anatomy, the reason why is just that a biological system evolved that way, and it being complex and highly particular is not a downside to the theory, it just means explanations for LFs origins have to be sought diachronously (like: maybe the capabilities of the LF and the actual emerged forms of language co-evolved, influencing each other). The other approach that I prefer because it is simpler is to explore how much of the presupposed rules of language can themselves be found to be part of one single, simpler generative / combinatorial process. This involves de-“reifying” (2) standard features of linguistic theories like “NPs”, “clauses”, the “argument structure” of verbs, and “phonotactic constraints” to show how these apparent regularities are not rigidly codified rules at the outset of a language’s description, a la axioms, but they are resultant patterns, sort of ad-hoc constructed themselves from a much simpler underlying constructive process, and they permit more variation than is realized because a problem in abductive reasoning (3) - a working hypothesis is suggested (like: if an argument of a verb/construction is not realized, then it is still implicitly available as a ‘trace’ in ‘deep structure’ (4) - there is maybe a minor chance of unfalsifiability like Russell’s teapot (a really interesting discussion related to that here (5)), and when issues with the working hypothesis are found, it is maybe easier to maintain some kind of meaningful order and control on at least a subset of the data by excluding exceptive cases and not rejecting the premises of the theory. Unfortunately, this is also a classic pitfall of twisting facts to suit theories, where it may have been forgotten that the theory may have only ever been tentative or conditional, to begin with.

In other words, this thinking/theorizing pattern:

“I am noticing a pattern” (‘l’ can be followed by ‘p’, but not ‘q’) -> I suggest a theoretical construct / concept as a tentative “rule” of some system ( - I don’t care about why that rule exists, that can come later), such as, “the LF is such that it selects permitted and unpermitted phoneme clusters as a hard rule that then determines the observed forms of words (I can suggest reasons from the physiology of the mouth, entropy, Zipf’s law, a need for sufficient contrast and ambiguity resolution, energy efficiency aka “economy”, and so on, but these are not taken as laws of the theory yet as the assertion of a phonotactic constraint law is, just accompanying thoughts) -> if I ever notice a subtle exception to the rule, I can claim it is not a true exception because it is actually “non-linguistic” (like, “intentionally un-grammatical expressions” like “how you do?” from a native speaker).

So doing, I am able to posit the existence of some rule-based system that has a decent capacity for evading falsification (in David Deutsch’s words, “people should trust theories that are hard to vary”).

What the above suggestion of a rigid law about allowed phonemic pairs might miss, is (I use this word all the time), it has reified sort of like, an instance of a pattern into.. a law rather than something transient, is the only way I can think to say it. An effect rather than a cause. Or maybe the difference between “happens to be so” vs. “and therefore must be so”, true vs. necessarily true (vs. necessarily necessary, vs. necessarily necessarily necessary…)

The important thing is that a rule will ultimately need another rule to explain where it came from - if you hope the entire universe is explainable and coherent, and science / understanding the world is a meaningful endeavor and possible.

When we separate different parts of the rules into different domains like “English grammar”, “the language faculty”, “the evolution of the language faculty”, “human cognition / psychology”, and ultimately even physics, mathematics, and eventually logic - there is a chance (even if in practice it is a good approach) that it makes it easier to a) observe a pattern (there appears to be some dynamic or tendency regarding heavy objects and how they move, I wonder if I can spot and guess through mental trial and error what the essential factors at work appear to be), b) codify (merely formulate, express) the pattern (like, F=ma), c) reify the pattern as a law (this can be the bad part - I noticed that force always equals mass times acceleration, because look how many examples in the world around me I can interpret as being in accordance with my law, particularly given my choices in how to categorize and conceptualize the thing I am observing); it must be a universal law of nature that this is so. I keep noticing the pattern, so I assume the pattern causes the data - d) don’t worry about where the pattern came from (How could I know? God made the universe. It’s just the way things turned out.)

The opposite is treating rules identically to the objects of study, where they necessarily need rules to explain where they came from, and ideally to abstract away the idea of boundaries between this field and that field (syntax, human cognition, the anatomy of the mouth, biological evolution) that allow you to do nothing but generalize a subset of some data as a kind of “parametrized” form (like, VP -> VP + NP “parametrizes” every specific verb by suggesting the same law applies to them all and you can plug in any verb you like to generate a specific construction rule on that specific verb), and commit the fallacy that “a possible rule/an observed rule is a necessary rule” (I think). It is a surface manifestation that there appear to be phonotactic restrictions characterizing languages, but the reason why can be so many different reasons or rules at the same time.

The idea is sort of, by not categorizing “rules” into different domains of knowledge or compartments of the physical world, we can just study the abstract rules, virtually, required to generate information of a certain distribution, and never exclude a needed part of a total explanation almost by delegating it ostensibly to a different field, for example, anaphora resolution: assuming humans (sometimes, but commonly enough to be beyond random, seemingly) end up figuring out what the referent of a pronoun is, the conceptual approach would ask how: some module of human cognition presumably performs this function - is it a simple statistical analyser like in “centering theory”? Or is it a deeper blackbox of human consciousness for linguists, for neuroscientists to examine? It actually doesn’t matter what the physical substrate of whatever is carrying out the computation is - it will still have to be a dynamical physical system that updates its states according to the laws of physics / our universe - so we don’t need to abstract the question away, as “non-“ or “extra-“ linguistic as opposed to “linguistic”. Then, the important question I think is, what computation class the patterns in the data appear to have. Ie, to exist in our world, it has to be “computable” in a sense - realizable, performable - but the question is, what general kind of a “process” is it, of what fundamental mathematical type? Because there are are mathematical objects and systems which classical computers are incapable of attaining - but I believe the actual physical universe appears to have systems embodying mathematical properties beyond Universal Turing machines - so if consciousness actually uses higher-order mathematics, we can consider how the “algorithms” of language are not realizable by like, a normal Turing machine, but there can be other logical descriptions of how they work.

I think this might be what model-theoretic linguistics does, but I’m not sure. Maybe something like, any observed pattern - any subset of linguistic data - can be framed locally in any mathematical context which could model that extracted phenomenon, that subset of data, but it would never imply that the data can only be explained with recourse to that one abstraction of the subsets’ specific patterns. I believe the strength of this approach is it would encourage always taking a higher level of abstraction on whatever patterns you’ve currently found, to frame them on an even higher level, so maybe, it is easier to fix only partially correct rules by framing or embedding them in a higher level rule which one can see has that former rule as a subcase or one of its (minor) consequences.

Anyone care to stimulate me on these points, guide me towards some strong ideas in the field, or work on this topic?

Answer 1

Step 1 is to say what theories there are of the nature of so-called phonotactic constraints. Phase 1 was Morpheme Structure Rules, which held that lexical items could be partially specified for certain features that can be supplied by rule, for example a morpheme that begins with /s/ plus a labial consonant either begins with /sm/ or /sp/, and other features can be filled in by rule. This theory was pursued in the late 50's in Halle's Sound Pattern of Russian. Phase 2 was adopted at the point of Chomsky & Halle Sound Pattern of English, where fill-in rules were replaced by "evaluative constraints" that operated on fully-specified lexical entries. There was a technical motivation, impelled by Richard Stanley's article on redundancy, that pointed out (presumed) over-generation problem with having three lexical values rather than two. The basic difference between the two approaches was that predictable features were actually removed from the lexicon in the earlier account, but were "present, yet not counted" in the SPE account. The common feature is that both theories assumes that the simplicity of a language was in part computed based on counting features in the lexicon. In this class of approaches, the more-predictable a property is, the simpler the representation, and the simpler the representation, the more likely it is for the representation to exist in a language. There is hand-waving in that premise, see below.

Phase 3 came with autosegmental phonology, which viewed representations as a network of association relations between features, as opposed to being a series of matrices with a fixed set of rows. The theory still adhered to the view that at the level of pronunciation, all segments have complete specifications of all features – it's just that that was not a requirement of lexical representations. So rules ("default rules") filled in anything that was missing. This is basically a reversion to Phase 1, which is now on a firmer theoretical basis (mainly, a carefully-curated disinterest in over-generation plus the empirical necessity of partial feature specifications). Phase 3 led to Phase 3.1 where features as binary attribute-value pairs gave way to bare attributes, e.g. [+nasal] become [nasal] and there is no [-nasal]. The consequence of that move was that no fill-in rules are required. This did leave phonotactics in a rather uncertain state.

Phase 4 started with the realization that there may be no bare phonotactics, there are rules of syllable-structure construction, starting with Kahn's theory of syllable construction and culminating in Clements & Keyser's CV phonology. Kahn simply made the sequencing constraints be part of the rules building syllables (or, claimed to, but he never actually formalized the rules), but CV phonology moved most of the equipment for building syllables into UG in the form of "constraints". This eventually became Phase 5, Optimality theory, where all grammatical statements are expressed as prohibitions, and the grammar is an ordering of those prohibitions (the list of prohibitions is itself part of UG).

Step 2 is addressing the question of whether there are any such devices in grammar. Now things get really complex. The simplified answer is that there are rules that regulate how syllables are built, and most "phonotactics" fall under that rubric. The other half of the answer is that "phonotactics" is the reification of "the result of applying rules". The standard rule-based approach to post-nasal voicing is to say "there is a rule voicing obstruents after a nasal", therefore a "constraint against [nt] is the reification of the fact that there is a rule /nt/ → [nd].

Ignoring the rest of the question "where is this in the grammar?", attention turned to "why are the facts this way". Over the past 20 years there has been an increased awareness that the theory of grammar needn't replicate theories of anatomy and physics, so if we can explain why the data are the way they are without imputing that mechanism to grammatical theory, then we will have learned something (mainly, that phonotactics is epiphenomenal).

There is, hence, a cottage industry of finding functional explanations for why the data are the way they are. (Children learn grammars based on the nature of the data, so language acquisition explains why grammars are the way they are given the data). There are a number of functional explanations for particular generalizations, for example the fact that English syllables can start with [tr] but not [rt] is based on the perceptual difficulty of perceiving [r] in initial [rt].

In "short", the best explanation is to take the problem out of grammar and make it be an issue of historical linguistics. Why do we not have syllables that start [nd, nt, mb, ng] in English? In part because they are a perceptually difficult so children are likely to miss the initial nasal in such syllables. But also in part because our ancestors didn't talk that way either, and that's a generalization that has remained mostly valid for many thousand of years in Indo-European (with exceptions, such as Sicilian). Bantu languages, on the other hand, have no problem with such onset clusters (but they do have a problem with [br, sp] etc.). Rather than asking "Why don't we have X?", we ask "Why would we have X?".