At the Bat: A Predictive Language Processing Primer
By: Cooper Howland
Step up to the plate
Imagine a major-league baseball game. Your favorite batter steps up to the plate. The pitcher rears back and throws. From the moment the ball leaves the pitcher’s hand to the moment it crosses the plate, approximately 500 milliseconds (ms) will elapse. That’s half a second. In that time, the batter must evaluate the pitch headed his way, adjust his stance (150ms), and swing the bat (200ms). He is left with 150ms—half the time it takes to blink his eyes—to determine what kind of pitch it is and to decide whether and how he should swing, or indeed if he should fling his body out of the way of a wild pitch.
He does not have time to consciously make decisions about whether he faces a fastball or a change-up. What he does have is a cognitive framework built over thousands of at-bats that allows him to make predictions about the pitch sequence in the context of tonight’s game, every game in his entire career, and everything he knows about the opposing players. He has, since childhood, developed an entire specialized neural structure dedicated to predicting possible outcomes of the pitcher-batter interaction and matching them to the tiny sliver of input his eyes receive before the bat must either be swung or held steady.
That’s nice, but wasn’t this supposed to be about language?
The brain, it has been claimed, is not a reactive organ, but a prediction machine. Predictive processing has been hypothesized to extend to all parts of human life, from operating a motor vehicle to experiencing a piece of art, and language is no exception. Researchers have further hypothesized that the brain of a person having a conversation behaves very much like the brain of our baseball batter.
The brain:
- uses past experience to make predictions about possible upcoming lexical, syntactic, and phonological events. (This pitcher tends to get wild around the sixth inning; be careful.)
- prepares possible responses to these hypothetical events. (Don’t swing unless it’s right down the middle. He’s likely to throw a ball. If the pitch looks decent though, swing hard.)
- compares them against the input from our sensory organs (It’s coming low and inside.), and
- commands our various organs and muscles to conduct the prepared response (or some version of it) that seems most appropriate to the input. (Don’t swing; it’s likely a ball.) Finally, it…
- returns to step one, revising its predictions to prepare for the potential impact of our linguistic output. (He just threw a ball; he’ll probably be more cautious on the next pitch. Get ready to swing.)
All this happens in real-time in a series of delightfully recursive loops, with very little conscious input from any of the participants in the conversation.
Errors
Let’s consider a hypothetical conversation using the model laid out above:
- Your boyfriend takes you to a nice restaurant for lunch. Things have not been so good between you two lately, but you know he has been trying to improve the situation. His facial expression is emotional, but you’re not certain what exactly he is feeling.
- It’s possible he’s going to propose. It’s also possible he’s going to break up with you. Your brain prepares an appropriate response to each.
- He reaches for something in his pocket, pauses, smiles sheepishly, then starts to speak. “Honey, I love you more than anything in the world. Would you…”
- He’s proposing! You go with your prepared answer, “Yes! Yes of course I will marry you!”
- He looks confused, then ashamed. “Oh, no…sorry. I forgot my wallet. I was going to ask you to pick up the bill.” You are, of course, speechless.
This situation, while extreme, seems an appropriate metaphor for the experience of the language learner when they try to interact in their L2. Everything is going along smoothly until the moment when the external input bears no resemblance to any of the predictions the brain has prepared. It is as though our batter is preparing for the next pitch and the pitcher instead heaves an enormous, multi-colored beach ball or perhaps a handful of emeralds at him.
At this point, there occurs a sort of neural firestorm that scientists have identified using EEG equipment and named the N400 event-related potential.
Think about any conversation you’ve had in your non-primary language, when everything was running along smoothly and then your interlocutor uses a word you’ve never encountered, or perhaps uses a known word in a way totally unfamiliar to you. Alternatively, consider the faces of your students as it becomes clear that they have not understood the instruction you just gave them. Conversation, meaning, communication grind quickly to a halt.
In that context, it is perhaps unsurprising that the L2 user’s brain has been shown to be considerably less efficient and effective at predicting the next word, phrase or sound than that of an L1 user. They have spent far less time building up the specialized neural structures necessary to use the L2. Their contextual framework contains fewer nodes. There is no doubt that L1 language users are better at predicting the content of their L1 than someone who is still learning it. The main debate at present seems to be whether it is possible to strengthen a language learner’s predictive abilities, or whether that just sort of happens as they gain proficiency with the language.
Shall we bother?
Huettig and Mani (2015) make a case that predictive language processing is not necessary to understand a language, referring to it instead as a “helping hand.” After all, a student of a language can sit alone with an L2 novel and their dictionary and eventually puzzle out most of the meaning word by word. You can put a foreign film on your laptop, subtitles off, and, with enough rewinding and Google Translate, figure out everything that the characters are talking about. But that doesn’t sound like much fun, and it certainly doesn’t sound like authentic interpersonal communication.
I would argue that the ability to predict upcoming utterances is necessary to use a second language in real time with anything resembling one’s fluency in the first. It is the thing that makes the difference between halting, floundering gridlock, and smooth, coherent communication. Perhaps more importantly, the ability to predict the general format of upcoming input allows the L2 user to focus on the content—the message that their interlocutor is trying to get across to them—rather than getting tripped up on confusing functional language. In that context, it seems that strengthening predictive ability is absolutely a worthwhile endeavor for the language educator.
How shall we go about bothering?
Curcic, Andringa and Kuiken (2019) exposed 150 subjects to a miniature “novel language” of their own invention. They then tested the abilities of the subjects to predict a determiner-noun agreement pattern. They found that if a subject was aware that the determiner predicted the noun, they were more likely to predict the upcoming noun, and to do so more quickly. This may not come as a surprise. One must be aware that a pattern exists before being able to recognize and exploit it. The really relevant part for us is Curcic et al.’s observation that students who received explicit instruction regarding the determiner-noun agreement pattern were less effective at predicting it than those who had worked it out for themselves.
To return one final time to our beleaguered baseball metaphor: it is marginally helpful for the coach to tell his batter, “this pitcher always throws a fastball on the first pitch.” But after the batter has personally faced that same pitcher a few times and seen those fastballs coming in like clockwork, he knows in his bones that the first pitch is going to be a heater. He may not even be able to put into words what he knows or how he knows it, but he knows just how he has to swing on that first pitch. He has moved beyond explicit knowledge, and into the realm of implicit knowledge.
In grammar teaching, this process is referred to as “noticing.” Call it instinct if you like, or reflex. Using the terms we are discussing in this paper, he is predicting accurately and consistently.
There is a huge difference between explicit instruction and pattern recognition, and plenty of controversy in our field about which one is more appropriate in a given setting. If one has decided to prioritize the strengthening of predictive skills, Curcic et al.’s research indicates that guided inductive teaching may be the way to go about it. That is to say, try not to think of a language as a set of rules and their exceptions that you, the educator can simply explain to the students and be done with. Rather, consider setting up the classroom as a place for the collaborative uncovering of mysteries, peeling back layers until truth is arrived at together. Help students to develop their own instinct for the language, their own “specialized neural structure,” their own ability to predict and respond to input on an instinctual, reflexive level.
Help them knock it out of the park.
Edward Cooper Howland lives and works in Hiroshima. He enjoys bringing research from outside disciplines to the ELT community’s attention. He has recently developed a fascination with the sport of curling.