Machine Learning Doesn’t Care About Grammar, So Why Should We?
Early machine learning practitioners
One of the strangest parts about teaching a machine learning model to perform speech-to-text is that you teach the model language very similar to how you would teach a child. You have them practice, a lot, as there is no fast route to learning a language, regardless of whether you’re teaching a four year old or a deep learning model nested somewhere in Google Headquarters. Both the kid and the machine learning algorithm will also require tons of time to make mistakes so that they can improve. And, with enough practice, both can become very, very good at understanding a language, whether it’s English or Icelandic.
However, to learn a language, neither the four year old nor the machine learning model needs to know a single thing about grammar to learn language. No verb conjugation charts, no sentence diagrams, not even a single explanation of “there” vs “their” vs “they’re”. Yet, in our approach to language we are obsessed with grammar and, in fact, it is the primary lens through which we teach the subject. Why is this? Why can four year olds, who are mostly pretty dumb (sorry!), learn a language without taking a single class, and yet we think that it’s impossible for an adult to learn a second language fluently? Well, part of the reason may be that we don’t teach second/third/fourth languages the way our brain wants to learn them. After all, adults actually can learn a new language fluently, the CIA trains people to do just that all the time. So why do we focus on grammar so much in the classroom, if it’s not getting us the language skills we really need?
We know that trying to build a language algorithm based on grammar doesn’t seem to work because this was our first approach in the 1950’s, and it was a gigantic failure. Before artificial intelligence was possible, people still tried to build computer programs that could handle language using “rule based” programming, where the rules were mostly grammatical. Verb conjugations, conditional clauses, and all the other stuff you learned in Spanish/French/whatever class in high school were programmed into an algorithm. The results, however, were uniformly terrible, also probably the result of your attempts to speak that second language after having taken that class in high school. Language has way too many exceptions to build a rule for everything, everything became too complicated and spiraled out of control too fast. It wasn’t until we started using machine learning, where we let the algorithms create their own rules, that speech-to-text and other natural language processing ideas started to get somewhere productive.
For linguists, grammar is the way we view language because, when we break language down these are the underlying structures we find and can describe. Breaking down language to get grammar is very similar tearing down a house and making observations on finding a foundation, drywall, and a frame — all critical components that help make a house what it is. And grammar is valuable! It has come a long way towards helping us understand what language is and how it works. But grammar is what we find when we break down language, but it does not mean that we can recreate language using only grammar. Much like we can’t tear down a house and instantly understand the nailing schedule or each and every circuit capacity, we can look at grammar but still not understand why every irregular verb is the way it is.
This means that using grammar to teach a language may turn out to be going in the wrong direction, much like it was when we took the grammatical approach towards building language algorithms. It feels a little like using geometry for “Lecturing birds how to fly”, as the Nassim Taleb says. You’re working backwards by assuming that you can recreate a system as complex as language as easily as you can break it down, when this is almost never actually the case. If you’re interested in how to learn a language more effectively, check out “Why Don’t We Teach Language How We Learn Language?”
- Jack Connor