Inputting a Language, or Training the Neural Network That Is Your Brain
If you hang around language learning forums long enough, you’ll hear the phrase “inputting a language.” Most of the time that you hear this phrase, the author will be emphasizing just how important language input is (myself included 🙋♂️). But what does “inputting” a language even mean?
Input is a peculiar word choice, because “input” connotes data entry on a computer. While learning a language is anything but a cold and mechanical process, “input” is an apt word because it reminds us of certain immutable truths about language learning.
Specifically, that your brain, while marvelous, does not learn by chance.
Like a sophisticated machine, your brain learns in a well-understood manner. And understanding how it learns can help you avoid going astray with your precious study time.
In fact, the recent breakthroughs in artificial intelligence or “machine learning” have taught us a lot about how we learn. So let’s dive into a little history and explanation of machine learning so you can figure out how to improve your own language learning.
How Do Machines Learn?
Experts used to think that the human brain’s intelligence was akin to knowing a bunch of rules. In this way, “learning” was about getting the rules.
Under this model, learning a language was all about mastering grammar rules and vocabulary, because communication simply required you to marshal words in logical sequences to convey meaning. This is a model still taught in school—that learning a language just requires you to drill grammar and vocabulary.
But this model of learning has been debunked (much to the Anki and textbook enthusiasts’ chagrin).
As Google first showed with its improvements in translation, and OpenAI has shown by creating AI chatbots with uncanny language abilities, language skills are more readily achieved through machine learning models that embrace trial and error rather than those that systematically apply rules.
These trial-and-error models are designed to imitate the structure of the human brain. It’s why they are called “neural” networks. And because neural networks learn a language in the way our brain does, we now have new insight into the fundamentals of language learning.
That is, trial and error is the key to learning a new language.
How Neural Networks Learn a Language Through Trial and Error
Neural networks learn a language by training on massively large data sets. These data sets are libraries of sentences of natural language that are input into the model to reduce its predictive error—that is, to make it process language naturally.
In fact, a neural network’s ability to naturally interact with a language is entirely dependent on the quantity and quality of example sentences the model inputs. This is why neural networks used to process language are also called LLMs or Large Language Models.
Emphasis on “large” here, as the sheer amount of training data required is staggering.
Because neural networks learn a language by inputting a massive amount of sentences, you too will only learn a language by “inputting” a large number of sentences. And the degree to which you will be able to interact with the language in a natural way depends on how many examples you input.
This means that inputting a massive amount of natural sentences is the only thing that matters for learning a language.
Input is the key to learning because it is how a neural network “learns” through trial and error.
Inputting has 4 steps: (1) Prediction, (2) Comparison, (3) Reflection, and (4) Adjustment.1 It looks like this:
- The model tries to guess the meaning of a sentence it inputs.
- The model compares its predictions against a true result and determines if it made a mistake.
- The model figures out what caused the source of its mistake.
- The model adjusts the way it guesses to avoid the mistake.
By repeating this process over and over and over again, LLMs eventually learn how to handle a vast array of natural language input and make accurate predictions (i.e., correctly guess what the input sentence means).
How to Input Like an LLM
Like a neural network, you need to go through 4 steps to input a language:
- Read or listen to a sentence and guess what it means.
- Compare your understanding against the true meaning (i.e., a reliable translation).
- Reflect on the source of your error (e.g., was it a word you didn’t know? A verb conjugation you were not familiar with?)
- Adjust your understanding (e.g., look up the word you didn’t know, review the verb conjugation you weren’t familiar with)
This leads to a few takeaways.
First, active reading and listening are essential: you are not inputting if you merely hear or look at example sentences in your target language. Learning comes by struggling to understand, checking your understanding, reflecting on and adjusting your understanding, and repeating this process over and over and over again.
Learning requires you to strive to understand each sentence you input. This means checking your understanding against a reliable translation as needed. This also means looking up words you don’t know and reviewing grammar points as needed. These steps are essential so you can adjust your understanding and improve.
Second, because the first step of inputting is attempting to understand a sentence, you must have at least a foundational understanding of a language’s building blocks to input effectively. For 日本語, this means knowing ひらがな, カタカナ, and a core of essential vocabulary and grammar.
Once you have this core knowledge, you will be able to unravel a sentence into separate parts, look up what you don’t know, and make an intelligible guess as to the meaning of what you are inputting.
Third, truly understanding how words or parts of speech are used may require many, many examples. LLMs don’t just pick up discrete facts about a language on the first go—they may still fail to accurately predict the meaning of a word or the nuance of a phrase even after training on many, many examples.
In 日本語, this is especially true for the use of particles like が and は, adverbs like あくまで, and words like 並ぶ that can have multiple meanings (i.e., “to line up” and “to rival”). Sometimes your understanding won’t become clear until you’ve seen many examples of what the expected outcome is. And that’s just the way things go with language learning.
Putting It All Together
Inputting a language takes time and can be frustrating, especially when you are just starting out. But you can take solace in knowing that input alone is the key to language mastery.
Remember that it’s only by inputting a vast number of natural examples that you will truly understand a language, including what sounds natural and what doesn’t.
So focusing on anything other than inputting sentences made by and for native speakers is a waste of time.
Studying grammar and vocabulary in isolation will not make you fluent. Drilling with textbooks is not going to get you there either.
To speak naturally, to understand natural language, you have to input natural language. So drop your training resources as soon as you have the essentials to start inputting.
And then get ready to input a lot!
- In LLM speak, this is (1) Forward Propagation, (2) Loss Calculation, (3) Backpropagation, and (4) Updating Weights. ↩︎