What Language Learning Cannot Be

I noticed that W. Stanley Jevons’s remarkably successful little book Elementary Lessons in Logic (reprinted annually for decades after its appearance in 1870) uses language learning to illustrate two ways of acquiring or transmitting knowledge (see Lesson XXIV, “On Method, Analysis and Synthesis”). One is the method of instruction:

A student, for example, in learning Latin, Greek, French, German, or any well-known language, receives a complete Grammar and Syntax setting forth the whole of the principles, rules and nature of the language. He receives these instructions, and takes them to be true on the authority of the teacher, or the writer of the book; and after rendering them familiar to his mind he has nothing to do but to combine and apply the rules in reading or composing the language.

If only. Those with a classical education may come to believe that this is how they learned Latin, but it is not so. In Jevons’s day no “complete Grammar and Syntax setting forth the whole of the principles, rules and nature of the language” was available for Latin or any other language, even English; and little has changed.

Grammars are never complete: Natural languages are so rich and diverse in variety of sentence types that there are always residual undescribed constructions. The Cambridge Grammar of the English Language (CGEL, 2002) seems to be the first published grammar ever to note the reduplication of modifier words for emphasis in a major, major problem, or much, much too far. CGEL is as comprehensive a reference grammar as has been compiled for any language, but there are phenomena it does not cover. One such is the way you can add the regular past participle suffix -ed to a noun and use it with the intransitive preposition out to get a way of describing the state of being sated with something: I’m all caffeined out; We decided we were all museumed out; etc. If you understood the examples in this paragraph, some of your learning was done entirely on your own.

Note also that the hapless student presented (implausibly) with a complete statement of the principles and rules governing a language would probably find “rendering them familiar to his mind” close to impossible. Stated explicitly, or even informally in plain English, the rules and principles are daunting to behold and near impossible to memorize or apply.

Jevons’s other mode of knowledge acquisition is the method of discovery. Applied to language, it comprises “a tedious comparison of letters, words, and phrases, such as shall disclose the more frequent combinations and forms in which they occur,” and it “depends to a great extent on the happy use of conjecture and hypothesis, which demands a certain skill and inventive ability.”

Again, conceivably those few who have explicitly puzzled out how to speak or read a language with no assistance from books or lessons may believe that this is what they did; but it isn’t. Certain talented linguists (Daniel Everett, featured in this 2012 Chronicle article, is one) are capable of demonstrating in front of an audience how a field worker can gain grammatical information about a previously unknown language from a native speaker without interlocutor or intermediary contact language, and can develop a rudimentary grasp of some basic grammar points within an hour or two. It’s known as “doing a monolingual.” And it’s not done by identifying “the more frequent combinations and forms in which they occur” through “a tedious comparison of letters, words, and phrases.” It is done through nonverbal interaction with the native speaker (eye contact, body language, signing, miming with props), plus ostension, imitation, common sense, guessing at meanings, trial-and-error testing, and above all, educated intuition about how human languages generally work.

The “tedious comparison” of letter, word, and phrase distributions that Jevons alludes to looks remarkably like a probabilistic version of the procedures set out in great detail by Zellig Harris (Noam Chomsky’s mentor) in his 1951 classic Methods in Structural Linguistics. No one ever followed Harris’s semantics-free, mechanical modes of analysis manually. Only in recent years have computational linguists begun trying to automate stochastic implementations of Harris’s methods and apply them to language. It has not led to efficient machine learning of human languages, and probably won’t.

The truth is that we don’t understand how humans acquire languages. It cannot be through the method of instruction: The instructional materials are not available and would probably be useless anyway. It cannot be through the method of discovery: The task is too hard and (if feasible at all) would take too long. You can see the whole of the last century’s work in linguistics and psycholinguistics, especially since 1965, as an exploration of how and why Jevons couldn’t be right, and trying to figure out what on earth could.