Genie Out of the Bottle

Acquiring or Learning a Generable Language

A speculative model of acquisition

March 5, 2013

Learning:

1. Reflex responses

This is the “building blocks” of motor control and “muscle memory”. We all have some native responses or reflexes, such as pulling your arm back after touching something hot (or even something potentially hot, e.g. warm water coming out of a cold tap), or an expletive outburst when hitting yourself with a hammer (worth noting that this is not a linguistic production, but much like Tourette’s).

Speech involves a lot of fine motor tasks that are not native at all. Trills, for example, are very hard to learn for non-natives. And the first step is to train these responses, so that you are able to produce them easily.

2. Eliciting responses with automatic cues

This step involves becoming able to elicit the response (e.g. a trill) readily, and then internalizing the cues. By internalizing, I mean not only being able to produce them at will, but more or less spontaneously, bypassing conscious effect. A good example would be slip of the tongue, or collocations (say, “We’ll pull a rabbit out of the fire, then burn that bridge when we come to it”).

It’s not necessary that you know about co-articulations or phonotactics; but your mind knows them implicitly. Just like how your brain manages breath automatically for speech or singing or shouting, you don’t have to think “I need to draw a deep breath now in order to boost the amplitude.”

To produce naturally-sounding speech, these responses must be as readily available as how your irises respond to lighting conditions, or how your cough where there’s smoke, or how you automatically extend your forearms when tripping forward.

3. Leaving it to Generative Grammar

This is part where I consider most of what’s happening is ’non-learned’. It’s something that your brain does for you. I’m still unsure on the question of the ‘critical period’ though.

Random notes:
Linguistics aside for a moment, if we view the mind as a composite machine (made up of interconnected modular parts) for regulating internal stimuli, somewhat in relation to external stimuli (the environment, where fitness is tested) — then our responses (e.g. production of speech, or human actions) are not solely or slavishly dependent on external conditions (nor that in combination with innate instincts), but directed by the mind and its generated internal states, which tend to be consistent with, but not deduced by, the environment.

The bottle is made of brass, but the genie is not.

Exaptation:

What your brain does is essentially using its native faculties to wire up the mapping and interfacing. By native faculty I mean something like you can look at twenty pictures of birds, and instantly grasp the abstract class, the concept “bird”, though those individual samples may be vastly diverse.

You have native access to off-the-shelf mental capabilities such as ontology (as taxonomy), comparisons, similarity, metaphors, generalisation, reduction, induction, and so on.

Once you go into a linguistic environment, and start interacting with the language (provided the basics, e.g. reflexes, hearing, perception are all there), your brain will start to automatically digest the ocean of inputs, and dynamically restructure the (implicit) mental model of the language, with your native faculties serving like factory processing lines (imagine a bunch of sophisticated robotic arms moving around).

This is where we mash up the reflexes with what is generative.

Shallow Processes:

What is left is what I consider the more trivial bits (yet what conventional learning has been focusing on, e.g. vocabulary), such as building up (to avoid the term ‘learning’) the lexicon.

I tend to consider it uninteresting, since a child can acquire almost what there is to acquire about grammar, then get the majority of the lexicon afterwards. The most impressive bits have already happened.

In other words, it’s the part anyone can do, relying on basic cognition.

The same goes for pragmatics, how to say something appropriate to the context; and much, much later orthography, which is hardly part of “language” (in the sense of the evolved mental faculty).

And that’s why it’s backwards to fixate on teaching lexicon through orthography.

August 9, 2013

There’s a long-standing objection to considering the use of language a ‘behaviour’. I want to discuss this idea and its implication to SLA.

From the first perspective, language cannot simply be just another behaviour, if we ask the question: “Is teaching a language as simple as modifying behaviour?”

Because so far it has become quite obvious, that the effort involved when adults try to learn a second language serves as good evidence, that language seems to be too creative and unbound, and too complex to be ever induced by Skinner-box conditioning. You can feed a rabbit cocaine or zap it mad, it still won’t have language. Whereas a typical child will acquire it anyway.

But from a second perspective, the notion (language is behaviour) is not completely unthinkable. If we consider the differences between a longer and a shorter sentence, between Shakespeare and a well-spoken six-year-old, immaterial; if we ignore the variability in content, and the inventiveness of use, then use itself is very much a type of activity that is universally carried out across the species — that it is, in this context, a behaviour — it has adaptive value, it has certain functions that solve various evolutionary problems faced by its users to varying degrees of success, and it’s subject to social and sexual selection.

From this point of view, in terms of how it relates to SLA, we see that the human mind has an instinct, a bias, a special liking, a kind of ‘sweet tooth’ for grammaticality, for rhythm, meter and rhyme, for poetic licence, for eloquence, for lyrical patterns and expressiveness — much like some birds having an innate preference for their nests to be round.

In summary, we cannot train people on content, we can only train people to use their faculties (to connect them to the systems of the target language) which then produce legitimate content. To borrow an analogy: teaching a student to memorise Plato is not teaching him to think.

If I were to start a school of thought in SLA, then it is this: instead of training students to memorise words and produce conversations, we should break ‘language’ down to its elementary components: rhythm, clusters, pattern detection, masking tolerance… to see through the fabric of the world and re-build the laws of physics from the fundamentals, and to design a completely novel sets of training tasks which look more cognitive rather than pragmatic in nature.

A theoretical framework

January 6, 2014

• A language is made up of smaller components of phonology, morphology, etc.
• These components form various levels of expressions (words, phrases, sentences) based on rules
• These rules may be generative, or they may be branching exceptions, or they may be idiomatic exceptions.
• Think of exceptions as rules that apply to a more specific subset. Even morphemes and their meaning can be considered as rules.
• The frequency of these rules in a natural language show a non-linear distribution
• The process of learning a language is to internalise these rules to a point of being able to interpret and generate expressions
• We learn rules based on Bayesian evidence. Exposure to comparative materials gradually builds a map of which rules are validated with what certainty.

Evidence Gathering vs. Iterative Reshaping

May 13, 2014

1.

When we try to use a generative model to capture a language, we think of it as a large collection of rules.

Examples of rules in Finnish are:

WHEN [Conditions…] => the Object will be in the genitive.
WHEN [Conditions…] => the Objects will be in the singular partitive.

You can think of these rules as one-per-line, item-by-item specifications of how the grammar should work. Though in a grammar book, they are often presented in a chapter-based format. See: Uusi Kielemme

Additional Note: Peter had later brought up a very good point, that these rules may not be discrete (in fact, very unlikely to be). They have a complex and somewhat unknown system of entailment, that the setting for one rule will pull strings tied to other rules.

2.

Then imagine for all possible natural languages, there is a universal template, which is an exhaustive list of all the possible rules, in key-value pairs. What if the ending is ’n’, what if the ending is a vowel, and so on.

When all these values are set, you have a copy of the language or a variant of it. And you can judge the grammaticality of an expression by checking them against the settings of that particular language.

3.

Though its nature is debatable, I’ve heard of many anecdotal remarks that Finnish is “harder” than English. What does this even mean?

Here I’d like to propose a way of looking at this issue: Finnish is harder, not because it has more or fewer rules, but because the rules are “less intuitive” — through another measure, we can say that the varieties of mistakes a naive learner can make in Finnish is greater than in English. (Yeah, it’s also debatable whether this is due to people have been pre-biased by English or other languages.)

• Let’s assume that when a learner tries a new language, he makes certain assumptions, for example:

“Let’s always use the nominative case!”

• Until he hits an incident that proves this wrong:

“ERROR!! It should be in the genitive.”

• Let’s assume there are some normative patterns of how humans form assumptions about these rules. (Well, you may say this sounds like quite a leap…) Think of it this way: a rule like “use the singular case every time the number is divisible by 3.” is less likely than “just use plurals for numbers greater than 1.”
• Simply put, some rules are more intuitive than others, and humans have an innate “feel” for it. They will presume the more likely rule, until it’s proven wrong in that language.

By this logic, English is easier because you get lucky more often. You don’t even know the rules, you were just guessing, and moving on based on assumptions. But you are lucky in this language, because most of the time, your guesses turn out to be identical with the actual rules.

Whereas in Finnish, you hit obstacles everywhere. It takes a long time for you to update all the rules. And because of the distribution of usage, and the different frequency or abundance of the various rules. What is at the tail end may take years to figure out. Cf. How plurals in some languages (say, Standard Arabic, I’m guessing) are not entirely mastered until late adolescence…

Whereas, someone learning English may have assumed early on: “Let’s just keep everything in the same case.” He may encounter a few problems with the pronouns, but then it’s all good. In German, he has to negotiate the 12 different kinds of ‘you’. In Finnish, well, he’s basically fucked.

4.

A side note is how this model fits into the “Poverty of Stimulus” debate. Natural acquisition is so fast, because the child does not have to statistically account for what constructions are more likely, which requires an endless amount of data, and you never obtain “assurance”. You don’t know that all swans are white until you find a black one.

But the generative model makes it possible, you start with a prototype of assumptions, then update them when you encounter an incident that proves the assumptions wrong. You need to see only a few, and sometimes just one instance to correct the model, while all this time it retains robustness in being able to keep functioning. You always have a “working version” to fall back on, and the minor details can be sorted out iteratively.

On Perceived Authenticity

May 13, 2014

I have long struggled with describing languages that feature some significant levels of diglossia. The grammatical signatures of register are not so drastically implacable in English. While in many other languages they are.

This is why when you search for an instance of proficient use, you can only find examples that are constructed in an almost entirely different style with its own set of rules. And if you apply these rules to building a translated sequence of meaning, it always sounds so odd and awkward. It fails at achieving authenticity, whatever that means.

You probably could not build the sentence “That is a red car.” and still sound eloquent and natural and intelligent in every language.

We have long held the notion that every possible idea should be expressible in every natural language. I agree, however I think by doing so a speaker has to modify the language itself into something else, a branching version that shares some features with the original, if you will.

So what are we talking about when we talk about the perceived authenticity of a constructed expression.

• We assume that the language has a set of rules, say Set R.
• These rules, modulated by statistical likelihood, determine how plausible an expression is: which goes from “totally okay”, to “odd but okay”, to “something is wrong here”.
• Let’s say that someone is learning this language. He tries to gradually figure out and internalise these rules. He acquires a model: Set R’.
• He uses R’ to generate an expression for an idea:

Idea –> [Rules] –> Expression.

• This process will be calibrated by some additional “annotative rules” which the learner has acquired along the way.
• Now when we gauge authenticity, we are in fact asking, how plausible it is, that this Expression is generated from the original Set R. Or, where does it land on the scale.

It should be expected that the more R’ resembles or manages to mimic R, the more successful that model will be in achieving authenticity.

De/Re-generative Grammar

October 20, 2014

This is the notion that instances of actual language use emanate from Platonian representations, of abstract models of the language, and go through a process of systematic decay, distortion, or realisation when they are produced.

Running speech is a de/re-generated product of idealised speech. Colloquial grammars are de/re-generated from formal models of expression. Hence it’s difficult for learners to faithfully reproduce authentic expressions through mere imitation, because the output itself is the product of a process of decay — they haven’t got the original, and they cannot let it decay the right way.

An analogy is apple juice, you get it from apples, that’s fairly simple. But to concoct an artificial flavour that tries to imitate apple juice, how?