You Don't Understand Comprehensible Input

[00:00:00]When we start learning a language, we bring in all these assumptions about what language is. And when we try to discern which language learning methods are best, well, that depends on what we think languages are made of. The classical mainstream model of language learning that we all start with is very simple. Languages are fundamentally made from two things, definitions and rules.

[00:00:19]So the best method to learn a language is the method that teaches us rules and definitions as efficiently as possible.

[00:00:26]Then we use this criteria when it comes to comprehensible input. And that leads us to say, well, comprehensible input isn't a good language learning method.

[00:00:33]It doesn't teach you definitions and rules fast enough. So therefore, comprehensible input is a slow, inefficient way to learn a language. But what if our assumptions are wrong? What if language learning isn't fundamentally built from definitions and rules? If the classical model is wrong, then before we judge any language learning method, we have to ask a more basic question. What is a language actually made of? if not definitions and rules. What I'm going to show you in this video is the core hypothesis of comprehensible input actually assumes a completely different model of what language learning actually is. And what we're going to see is that this model explains native like fluency far better than the classical model does. The most succinct definition of comprehensible input is this. We acquire a language by understanding messages in that language. But notice the word messages [music] doesn't appear anywhere in the classical model. The classical model gives us definitions and rules. So what exactly is a message? Well, let's watch the exact demonstration Steven Crashen gave when he introduced comprehensible input to the world in the 1980s.

[00:01:47]Good cop. And here I'll draw a picture now.

[00:01:53]Cop is good. Yeah, sure.

[00:01:58]Cop.

[00:02:00]He's not giving you a definition of the word hand. He's not translating a sentence. He's not explaining the rules of German. Yet, when you watch this clip, he's still able to transmit a message. Even though you don't know German, you still somehow understand what he's saying. In a game of charades, communication is completely nonverbal.

[00:02:19]Yet the person gesturing to you can still somehow send you a message. So if a message doesn't even necessarily require words, what are we understanding when we understand a message? We're understanding how pieces of experience, sounds, gestures, actions, objects relate to one another through associations. When Crashen points to his hand while speaking German, the message makes the relationship between the language and what you're seeing clear enough for your brain to begin connecting them. This isn't passive spongel-like absorption. You don't learn the language simply because it's floating around you. This is active participation in the recognition of meaning. And the interesting thing is the idea that we learn languages this way just completely breaks the classical model. It requires us to create a completely new model to even have vocabulary about what we're doing when we're watching comprehensible input.

[00:03:11]What are we really doing if not learning definitions and rules? We're going to call this new model the associative model. And at the most basic level, the basing building blocks are experience linked together by association. An experience is the smallest unit of conscious awareness that can be distinguished by the learner at any particular moment. An association is a weighted directional link between two experiences where activating the first raises the probability that the second activates as well. The classical model says that your understanding of a language is a collection of definitions and rules you've learned to apply automatically. The associative model says your understanding is a weighted network built from experience. And what we call definitions and grammar are patterns that emerge inside that network. Now you might be saying, well, all that sounds very abstract. How are you actually going to prove that the associative model is better than the classical model? Well, we can show that it better explains the experience of native like fluency. And luckily, all of us already have something inside our heads we can examine to find clues about how native language fluency actually works. And if you haven't guessed, that's your native language. We're going to be examining our native languages.

[00:04:21]So, instead of just saying, well, kids learn this way, therefore we should.

[00:04:24]We're only going to look at how we experience our native language right now and show that what we experience is best explained by experience and associations as opposed to definitions and rules. I'm going to demonstrate four phenomena in your native language and by the end I think it will be obvious which model explains them better. Phenomenon number one, you can't define words that you know very well. Try to define the word love. Go into your head, locate the word love, crack it open, and tell me the definition.

[00:04:56]If you notice, when you reach for what love means, you don't find one clean definition sitting in a drawer. And that should bother us if words are supposed to be stored definitions. If a word were truly an atomic definition, then defining it would be the easiest thing in the world. You'd crack it open in your head and you'd read off what's inside. But that's not what happens. It seems instead that you don't have immediate access to a definition at all.

[00:05:22]To produce one, you have to start reaching for experiences tied to the word love. Maybe you start thinking about feelings associated with love. Or maybe you look at relationships in which one person said to the other, "I love you." Either in your own life or maybe in a movie or a story. Your brain doesn't reach for some atomic meaning first. It reaches for experiences, feelings, situations, relationships. And notice how strange that is. You understand love deeply. You recognize it across completely different situations and you use the word very naturally. But when I ask you to produce the definition that supposedly constitutes your understanding, you can't simply read it out of your mind. And that's because the classical model has the order backwards.

[00:06:02]It imagines that the definition comes first and that allows us to understand each use of the word. But our experience runs in the opposite direction. We encounter the word in thousands of situations. it becomes connected to different people, feelings, actions, etc. Then if someone forces us to define it, we try to construct a definition from that enormous collection of relationships. In other words, we don't understand the use cases because we understand a definition. We're able to produce something resembling a definition because we understand the use cases. And the associative model explains that very naturally. Love is not an atomic definition floating around in our head. It's a dense cluster of experiences connected by associations.

[00:06:43]Remove every feeling, relationship, experience, sentence, and concept love has ever been connected to. And it's difficult to imagine what meaning would actually remain. So, definitions are real and useful, but they're not primitive categories that explain how we understand words. A definition is just a compressed description of related experiences. Phenomenon number two, words, you know, seem to pull other words. I'm going to put a few phrases on the screen with the last word missing.

[00:07:12]Each of those words pulls another word from your mind without any conscious thought. Again, that's a strange phenomenon. For instance, why does jelly immediately get pulled from your brain here as opposed to any of the other tens of thousands of words you know? Why didn't your brain hand you screwdriver or television? The definition of peanut butter doesn't contain jelly. Peanut butter is a spread made from ground peanuts. Jelly is a sweet spread made from fruit. Nothing in either definition explains why activating the first should make the second easier to access.

[00:07:43]There's also no grammar rule that connects jelly or peanut butter. So again, very strange in the classical model. But in the associative model, this is exactly what we'd expect. You've encountered peanut butter and jelly together so many times that hearing peanut butter raises the probability that jelly gets activated. Or in other words, peanut butter is heavily associated with jelly. And this seems to describe our experience of real-time comprehension very well in our native language. When somebody speaks to you in your native language, your mind isn't passively sitting there waiting for each new word to arrive so it can look up its definition. Everything you've already understood is constantly activating things that might come next. But individual word associations are only the smallest version of this phenomenon.

[00:08:28]We're going to come back to this idea later and show just how powerful this process becomes. But to get us there, we have to talk about the third phenomenon.

[00:08:37]We recognize categories that we never consciously learned. Let's play a game of one of these words is not like the others.

[00:09:03]Why can we recognize the outliers in these categories unconsciously? Why don't you first have to reason about what these categories are? In the classical model, we might expect that reasoning would look like this. Okay, red, green, and blue, those are colors.

[00:09:18]Ocean is not a color. Therefore, ocean is wrong. But instead, it seems like you don't even have to know what the category is to spot the outlier. If words are really just atomic definitions floating around in our heads, why can we recognize them unconsciously in a practically uncountable number of categories? Here's the answer. There's no reason in the associative model to imagine that words are the final stopping point of emergent concepts.

[00:09:44]We've already said that a collection of experiences can become so densely interconnected that we experience the entire cluster as one singular concept, a word. But once that cluster begins behaving like a concept, it can enter into new associations of its own. And when several concepts repeatedly share the same associations, those overlapping relationships begin forming a larger pattern. Take red, green, and blue. Each word already has its own network of meaning. But those networks overlap. All three are connected to vision, light, appearance, and the question, what color is it? Every time they occupy these same relationships, that shared pattern becomes stronger. Eventually, the overlap becomes stable enough to behave like one concept of its own, color. This is what we might call a high order concept, a pattern of associations that become stable enough to start acting like a concept on its own. And our brains seem to be able to create these high order concepts completely unconsciously in a practically uncountable number of ways. If this is getting a little technically dense, I want to give you an analogy for this idea so you have the intuition behind what we're showing here. Think about a melody. At the simplest level, one note can be associated with another note that often follows it. But when you hear a familiar melody, you don't experience it as one disconnected chain of individual note tootee associations. You recognize the melody as one larger shape.

[00:11:12]That larger shape is built from relationships between individual notes.

[00:11:16]But once the pattern becomes familiar enough, the entire pattern begins behaving like a single concept on its own. It begins behaving as a high order concept.

[00:11:28]For instance, I can change the key of this melody and you still experience it as the exact same melody.

[00:11:37]That means I changed every note, but you still recognize the high order structure. I can change the octave, making the pitch higher, and you still recognize the melody.

[00:11:47]I can add a variation, and you recognize it. [music] I can even change the mode, and if the melody is familiar enough, you can still recognize it. [music] The underlying pieces are completely different while the larger pattern somehow remains recognizable. And notice that we didn't have to introduce any new primitive category to music called melody. We started with individual notes played in relationships and those relationships themselves began behaving as concepts. The melody is real. You can recognize it. You can remember it. You can compare it to other melodies. But it's not an additional category sitting underneath the notes. It only emerges from the way notes relate to one another. And now I want to bring that same idea back to language. Two of these sets we started with were grammatical categories. Dog, table, and problem don't mean similar things, but they've repeatedly appeared in many of the same relationships inside English. Those shared relationships allow the entire cluster to begin behaving like one high order concept. what we consciously call a noun quickly feels immediately like an outlier because it isn't included in that high order concept. Him, her, and them have repeatedly appeared in another shared set of relationships. You don't need to know the name of that grammatical category to recognize that they doesn't belong. And notice what determines membership in these categories. There doesn't seem to be a rule stored somewhere deciding that dog is a noun and quickly isn't. Membership isn't rule-based. is simply the result of shared associations. The weird thing is the classical model makes this arbitrary distinction between certain high order categories. The concept of a color doesn't require us to store any grammar rule apparently. But for some reason, the concept of noun does, but I can make really unintuitive sentences that we immediately know are nonsense using colors. I'm going to the green today. I always know what blue it is. I never gave it to the red. Why would the set of colors require no rule-based system to verify their correctness while nouns for some reason do? Just because one high order concept is structural doesn't make it fundamentally different from every other high order concept our brain can unconsciously create. Or in other words, it doesn't seem like we need to posit some separate rule-based system to explain grammatical categories. And once high order concepts like nouns and verbs exist, they can form associations with one another just like words do. In this set of phrases, the individual nouns and verbs change, but the relationship between the two high order concepts remain recognizable.

[00:14:24]This relationship is called predication.

[00:14:26]That gives us grammar in the associative model. Grammatical categories are just high order concepts. And grammatical structures are just associations between those high order concepts. In the classical model, grammar is thought of as actual rules you store in your head.

[00:14:41]And you supposedly understand grammar when you can apply these rules implicitly, really fast, lightning fast application. But in the associative model, rules aren't things that your mind stores at all. They're descriptions of patterns that emerge through associations. So when we drill a rule to learn it, well, we're assuming that what our mind unconsciously stores is actually a rule. What we call a rule seems better explained as an emergent pattern built from countless associations. Just like a melody emerges naturally from the relationship between notes, grammar seems to emerge naturally from the relationship between words.

[00:15:15]Once again, the associative model explains our experience without us having to posit a separate system of implicit rules. Now, what we're going to do is bring everything we've demonstrated together to explain one of the most powerful features of native light comprehension. Number four, we can understand incomplete messages completely unconsciously. When you hear peanut butter, it activates jelly. But now we've shown that this process doesn't stop at individual words. High order concepts can activate other high order concepts. A noun-like concept in a subject position can activate a verblike concept. So at every moment in a sentence, everything you've activated so far is activating a weighted collection of words, meanings, categories, and structures that could plausibly come next. We'll call this collection your implicit prediction possibilities. Now, these aren't conscious guesses. You don't sit there listing every single possible continuation when someone's speaking. The possibilities are simply active with some carrying more weight than others. To show what I mean, let's look at this incomplete sentence. He blanked down on one blank, pulled a blank out of his blank, and asked her to blank him. Why in the hell, using the classical model, would you ever be able to understand this incomplete sentence completely unconsciously? The only thing I can think of that explains why you understand this is high order association activation or implicit prediction possibilities. Down on one activates knee. What gets pulled out of something when someone is down on one knee? What does someone pulling a ring out of their pocket on one knee? Ask a woman. Every word, grammar, concept, structure, and probably countless other associations chip in to illuminate the high order concept, a marriage proposal.

[00:16:59]And the moment we unlock that high order concept, the sentence doesn't even feel incomplete. Every single layer is operating at once. Individual words activate related words. Grammatical concepts activate other grammatical concepts. And the relationships across the entire message activate a larger situation that narrows every blank. By the time each missing word is supposed to arrive, your mind has already done the majority of the work for you. Each word clearly isn't being decoded from scratch. It's landing inside a network in which it's already been activated through associations. That's what makes native like comprehension feel fluid.

[00:17:36]Every piece of the language prepares you for the next piece. The incoming word is usually only confirming or adjusting a pattern that's already running. The classical model can tell you what knee, ring, pocket, and merry mean. Grammar rules could explain why we can anticipate the grammar. But neither definitions nor rules explain why you can remove most of the sentence including the grammar and still understand it completely unconsciously.

[00:18:00]And this reveals the real challenge for the classical model. If native like fluency depends on this network of associations, then any method that tries to build native like fluency through deliberate study has to somehow recreate this network. This creates what I'll call the hard problem of study.

[00:18:17]Nativeike fluency doesn't only depend on knowing what theoretically could come next. It depends on having an unconscious sense of what usually does come next and how much more likely one possibility is than another. Some associations are central and immediate.

[00:18:33]Others are weak, unusual, or only activated in very specific contexts. And those differences in strength seem to matter constantly during real-time comprehension. If you tried to construct that network through only conscious study, you'd have to decide not only which associations to teach, but also how strongly each one should be weighted, which examples should appear most often, which meanings are central, which uses are rare, which words strongly predict one another in what direction. But when you're learning a language, that waiting system is exactly the implicit knowledge you're trying to acquire. That's the unknown. That's what you don't know. That's the hard problem of study. To consciously construct a native like network, you'd already need the native like network to tell you which associations to add and how strong to make them. You'd have to consciously choose and weight the relationship between every word, every phrase, every concept, every grammatical structure, every situation, and every possible continuation. Nobody can build that network one deliberate association at a time. Most of that information is something that native speakers couldn't even explain. None of us know the probability with which one word or structure should activate some other word or structure. We simply feel what usually fits. Input solves the hard problem by removing the need to choose.

[00:19:52]The language itself supplies the examples, the frequency, context, and relationships to find the weights. This is why study can efficiently teach you what theoretically could be correct while still failing to build a nativeike sense of what usually is correct. And now we can finally step back and look at everything we found. We couldn't define love because the meaning wasn't sitting inside our head as one atomic definition. It was distributed across an enormous collection of relationships.

[00:20:18]Peanut butter pulled jelly because concepts don't sit in isolation.

[00:20:22]Activating one concept prepares other concepts associated with it. Grammar didn't require us to add a new category of implicit rules. Lower order associations formed high order concepts like nouns and verbs. And associations between those high order concepts produced grammatical structure. All of those layers worked together to explain how you could understand the marriage proposal sentence with most of the language missing. Your mind wasn't decoding definitions and applying rules one at a time. The entire network was active at once preparing the meaning before it arrived. So just by examining our native language, we can show how the classical model fails to account for native like fluency. We don't have to argue over how kids learn or squabble about studies we didn't read. We can just say that native like fluency doesn't even seem to theoretically be even possible to emerge from the classical model. The assumption that traditional language learners walk around with is this assumption that drilling a grammar rule until it's automatic is equivalent to understanding grammar. and the idea that fast recall of a word's definition is equivalent to understanding a word. But yet, when we take those ideas seriously, they face plant. They just don't hold up under scrutiny. Grammar doesn't seem like implicit rule application. Words don't seem like atomic definitions. And if you try to learn a language with these misconceptions, you will build your understanding of your second language in a completely unnatural way that doesn't concur with reality. You can build something, sure, but whatever that something is will be far removed from that native like fluency that you experience in your first language. And let me just bring my experience into the picture here. I've learned a second language and it feels remarkably similar to my first language, which suggests to me that there's no special mechanism required for building the second language that's just fundamentally different than the first. Now, of course, I can't fit the whole world into this video, so nowhere in this video did I really get in to how comprehensible input efficiently builds these associations. If you're asking yourself about how these associations actually get built efficiently with comprehensible input, I'm going to put a video on the screen right there that uh gets into that. And I will see you there.