Generative Linguistics, Recursion, Etc.

Ari Rappoport

At present (2018), most undergraduate teaching in linguistics and a substantial part of linguistics research involves the so-called `generative linguistics' (GL) doctrine. Most if not all of the central tenets of this doctrine are strongly disputed by leading linguists and cognitive scientists (e.g., Basic Linguistic Theory [Dixon, 2005], Cognitive Grammar [Langacker, 2008], Construction approaches [Goldberg, 2006], the Parallel Architecture [Jackendoff, 2002], a view from child language acquisition [Tomasello, 2003], a view from the uniqueness of language [Pinker and Jackendoff, 2005]). Some of the tenets have been abandoned by GL supporters. I devote a few paragraphs to GL in my R17 brain theory because its tenets still appear as truths in linguistics textbooks and because some of them are widely known to the general public.

The R17 brain theory implies that virtually all of the basic GL tenets stated during the last decades are completely wrong.

  1. There is no `innate' capacity specific for syntax, because syntax uses the general Q process architecture of the brain.
  2. There is no `poverty of the stimulus', since generalizations over simpler representations acquired through comprehensible input can be done just as they are done in all other modalities.
  3. There are no specific syntactic universals, because there are many types of auditory sequences that can be automated and that satisfy the motivational needs.
  4. The notion of `transformations' has no central role in language and probably stems from an English-centric analysis.
  5. The distinction between `performance' and `competence' is misleading, because every utterance that a person makes is a direct result of their brain representations at the time of speaking, and yields the update of these representations.
  6. Basing a theory on categorial measures of `correctness' ignores the facts that acceptability is a continuous spectrum and that the degree of perceived correctness depends on literacy, and contradicts mechanisms of language change.
  7. The large expressive power of language is due to the huge number of possible word sequences, which is in turn due to the large number of words and not to any inherent qualitative uniqueness of language.
  8. The brain does not implement recursion, as shown in some detail below.
There is basically no experimental evidence that supports these GL tenets, and much evidence that contradicts them (see references above). Other GL tenets include the obvious observation (already made by ancient Greek grammarians) that language is hierarchical (words, phrases, clauses), and the trivial observation that its significant event is the appearance of a new word (`the minimalist program').

One of the tenets of generative linguistics is that one of the things (or the thing) unique to language is recursion. Recursion means that there is an action operating on some input, and during its execution the same action is activated on some other input (usually derived from the original input). The execution of actions can obviously involve the execution of other actions; the point in recursion is that it is the same action that is being activated (with different inputs, otherwise we get into an infinite loop).

The claim to fame of recursion in language is that it supports an unlimited number of embedded clauses, for example in sentences such as `X said that Y said that Z said that...'. However, people very quickly (after two or three recursion levels) lose the meaning of such sentences; understanding deep recursive sentences involves a conscious mental effort and usually cannot be done without slow rehearsal. After just a few recursion levels, the only way to understand such sentences is to write them down, showing that the brain finds it particularly hard to implement recursion. This is a simple result of the Q process (the execution process presented by the R17 brain theory), because recursion requires the same node to maintain an arbitrary number of different open goals, which is exactly the situation that imposes strict limitations on the number of items in working memory (see WM section in the R17 paper). Indeed, converging evidence supports the proposal that automated language use (syntax, semantics) relies on a specialized WM-like capacity [Caplan and Waters, 1999].

The brain cannot do recursion: psycholinguistic evidence

There is solid experimental support for the argument above from the `missing VP' paradigm. Consider the following two sentences. Which one seems more natural immediately during listening (i.e., before re-thinking about their meanings)?

"The patient who the nurse who the clinic hired met jack".

"The patient who the nurse who the clinic hired admitted met jack".

It has been repeatedly shown (e.g., in English [Gibson and Thomas, 1999] and French [Gimenes et al., 2009]) that (1) sounds more natural than (2), even though it is syntactically incorrect. Syntactic predictions for the 2nd verb in (2) (`admitted') are created by the syntax node triggered by `the nurse'. The phrase `who the nurse' is followed by `who the clinic', which, since English does not use two consecutive `who' patterns, is represented using the same syntactic action nodes. Hence, this is a case where recursion should be used if it is used at all. However, the appearance of `admitted' sounds strange, and a plausible account for this is that the predictions (goals) for it have been overwritten after `who the clinic' was complemented by `hired'. Sentence (1) sounds natural because all words are from the same general semantic domain (so create anticipations that conform to the input words), while no syntax nodes are left with unanswered goals.

Note that the above does not imply that languages cannot use such patterns. It only implies that they cannot use recursion (i.e., the same node) to represent them. Indeed, embedding a clause within another is very common, and this is very simple to implement by each level being suppoted by different syntax nodes. A sequence of three consecutive main (non-auxiliary) verbs is not used in English, but it exists in Dutch and German, and the experiment above indeed yields less convincing results in these languages [Frank et al., 2015].