Views on Human Language

The following is written for other specialists in linguistics. It lists some points where my views differ from the generally accepted generativist views.

Morphology

The primary, preferred cognitive strategy in the processing of morphology is to look up memorized forms. Composition/decomposition is available as a backoff strategy but is dispreferred. A speaker is more likely to look up a chunked form walked than to compose it anew from walk + ed each time. The more frequent the form, the more likely that the speaker will do a lookup rather than compute the form. This is why irregular forms are more common in high-frequency paradigms.

The computing of novel forms (wug → wugs) may be more a matter of accessing and computing over known prototypes which are similar in ways such as rhyming (bugs, rugs, hugs) than of computing from scratch by applying ordered rules to an underlying form.

Many phenomena in morphology are properly explained in terms of the history of the language rather than as part of the synchronic competence of the speaker. The speaker memorizes whatever forms he or she is presented with. Linguists are able to recognize that the observed forms are the outcome of an ordered sequence of processes, but the processes in question are often historical and should not be mistaken for synchronic cognitive processes. The synchronic processes of morphology and phonology do not recapitulate the history of the language.

If a rule has exceptions (leaf/leaves), then it is a historical rule rather than a synchronic rule. There are no diacritic features; there is no partitioning of the lexicon.

Bracketing paradoxes are an artifact of the assumption that words are synchronically computed from morphemes rather than stored as a whole.

When confronted with a word for the first time, a speaker might be able to obtain some useful hints about the meaning of the word by noticing meaningful chunks within the word. The speaker who is able to segment constitution into constitute and -tion is consciously doing etymology. Once the word has been learned, the meaning is stored as if the word were monomorphemic, which is why its semantics need not be compositional.

A speaker who recognizes boy within boycott, and produces a novel form girlcott, is doing the same kind of processing as the speaker who recognizes constitute within constitution. Morphological processing is not concerned with etymological validity. The act of looking for meaningful chunks within words need not result in a complete or tidy morphological decomposition. The speaker can recognize that there is some sort of connection between divine and divinity based on general resemblance, but there are no synchronic phonological rules in Modern English which derive the two from a single underlying form.

Conscious manipulation of morphology is not the same as unconscious manipulation. Literacy and education influence the ability of the speaker to consciously manipulate morphology. Laboratory creations by linguists (e.g. anti-anti-anti-missile-missile-missile-missile) tell us something about the sorts of conscious manipulations which a learned speaker is capable of carrying out, but studying this kind of linguistic competence is like studying the ability of a mathematician to manipulate symbols when solving an equation. It is not useful as a way of studying the structures and processes underlying naturally occurring language.

Phonology

In general, explanations in terms of the constraints or biases imposed by the articulators and auditory system should be preferred over explanations in terms of innate cognitive apparatus ("universal grammar").

The cognitive component responsible for phonological processing is a general information processing system which picks up on whatever regularities it can find, without regard to the phonetic naturalness of those regularities. The fact that so many phonological rules seem to make some kind of phonetic sense is due to pressures from the articulators and auditory system. Palatalization is more common than anti-palatalization because of an anatomical tendency, not because an innate feature geometry makes it easier to represent palatalization rules. There is no reason to assume that the brain is innately, redundantly encoded with the same preferences which the articulators would impose anyway.

A phonological system is a product of its history. If something is rare or unobserved, it may be because there is no likely sequence of historical events which would give rise to it. There may be things which are cognitively permissible but which history does not tend to produce.

Why are there no plaid rocks?

Striped rocks are common, but plaid rocks are very rare. Why?

There is nothing wrong with plaid rocks. If you had millions of years to perform the experiment, there is nothing to stop you from piling up sand in the right way to produce plaid rocks. There is no inherent property in sand or rocks which prohibits plaidness. The molecules which make up the rocks do not come encoded with a "universal grammar of stone" which prefers stripes over plaids.

The reason that striped rocks are common and plaid rocks are rare is that natural processes, governed by very broad physical principles such as gravity, tend to deposit silt in horizontal layers, not in plaid patterns.

The same sort of thing is probably true of languages. The set of language types which we observe is the set of language types which history is capable of producing. There may be language types which are permitted by the cognitive apparatus, but which are never observed, because the processes of history tend not to produce them.

Syntax

Most processing of syntax is a matter of manipulating memorized chunks whose semantics are stored rather than computed. A chunk may contain multiple words. It may be an entire sentence; much syntactic processing may be a matter of substituting lexical items or adjoining phrases into stored prototype sentences. Composition/decomposition from atomic units is available as a backoff strategy, but is dispreferred.

Idioms are an instance of a more general strategy. A syntactic chunk is stored with a collective meaning. In cases where the stored meaning differs from the meaning which would be computed from the atomic pieces, we call the structure an idiom. However, it is also perfectly permissible for the stored meaning and the compositional meaning to be identical; this is probably true for the great majority of stored syntactic chunks. There are also chunks in the middle of the continuum whose meaning is not altogether idiomatic but also not entirely compositional (by hand, on foot).

A stored chunk does not necessarily need to be derivable from general principles. The sentence Long live the king cannot be fully decomposed according the the syntax of modern English, but speakers have no difficulty substituting a different noun phrase into this prototype sentence. Similarly, the English construction exemplified by the more, the merrier is highly productive, but does not fall out from any general account of syntax that I know of. The speaker simply memorizes the chunk and does substitutions into it.

A stored syntactic chunk might also be bundled with a stored phonological representation. A speaker might pronounce the first part of "Couldn't you have at least tried?" as /kʊdn̩ʧʌv/. As far as I know, the reductions in this sequence are not general processes of English; more likely, the speaker is retrieving a stored blob of phonological material. Speakers are capable of miswriting "could have" as "could of" because they are accessing a stored chunk which has a collective meaning and collective pronunciation, rather than atomically parsing the sequence. The higher frequency the sequence, the more likely it is to be stored as a chunk. Correspondingly, the higher frequency the sequence, the more likely it is to be stored with quirky phonological reductions.

A theory of syntax which allows center embedding to a maximum depth of three would probably suffice for the entire observed body of naturally occurring sentences. Most theories of syntax treat center embedding and right-branching recursion as mathematically equivalent, but there is obviously some asymmetry between them; center embedding is very hard and never goes very deep, but right-branching recursion is easy even at many levels of depth. The theory should account for this difference. An appeal to "processing difficulty" should spell out the specific nature of the difficulty, and should also explain why this limitation should be kept external to the model of syntactic competence, rather than built into the model of syntax itself. By comparison, it would not be particularly satisfactory to claim that universal grammar actually allows subjacency violations, and that the non-occurrence of such violations is due to some unspecified type of "processing difficulty".

Some techniques which have been worked out as engineering solutions to natural language processing problems, such as n-gram models, should also be considered as hypotheses in the human cognition of language. Humans may employ multiple parallel strategies in the processing of language.

Grammaticality is a gradient property rather than a discrete one.

Introspective grammaticality judgments are unreliable as evidence.