Hearing a baby’s first words is a joyful moment for many parents. But another crucial language milestone is harder to pinpoint for both parents and scholars of human development. When does a child start putting together words on their own, rather than parroting what they’ve heard?
A new study published last week in PNAS by researchers at the University of Chicago and others used behavioral and computational data to determine when English-speaking children go beyond their linguistic input. For linguists, this happens when a child uses a language rule to say something new—something they’ve never heard before.
The problem: it’s almost impossible to know everything a child has ever heard. To address this, the research team of linguists, developmental psychologists and computational analysts joined forces. They built a generative computer model that mimicked how a child first produces a certain structure in English: determiner-noun combinations (e.g., saying a dog after having heard the dog).
“We pinpointed the moment when we thought each child can do this, and then we tried to model that with a computer,” said corresponding author Susan Goldin-Meadow, the Beardsley Ruml Distinguished Service Professor in the Departments of Psychology and Comparative Human Development at the University of Chicago. “They agreed pretty well.”
Both datasets estimated that children begin producing determiner-noun combinations they've never heard at around 30 months. According to Goldin-Meadow, this novel approach, combining computational modeling with behavioral observations, opens new avenues to explore long-standing questions about how children learn language.
Learning from mistakes
We all learn by making mistakes. Looking for errors is also a useful method for linguists to assess how children pick up language. When a child says, “I eated my dinner” or “I thinked about it,” it means they understand a basic grammar rule in English: verb plus -ed means something happened in the past. Because English has irregular verbs, it’s easy to spot when a child uses this rule to produce a phrase they’ve likely never heard before.
For this study, the research team looked at a similarly characteristic part of English grammar: determiners, or words that modify nouns, like “a” and “the.” For example, a dog or the house. Researchers assumed that if a child used both “a” and “the” for the same noun, i.e. “a pineapple” and “the pineapple,” they likely understood the pattern and were using it to create novel combinations.
For the behavioral part of the study, researchers observed 64 English-speaking children and their caregivers. For 90 minutes every four months, they recorded parents interacting with their children and compared each child’s utterances to their parent’s utterances.
Based on these samples, they determined that children started using “a” and “the” in front of the same noun around 30 months. After their first instance, researchers also noticed that the children began creating even more combinations that weren’t recorded from their caregivers.
But a sample can’t account for everything a child has heard. “The children are sitting around listening to their parents every single day, but we aren’t,” Goldin-Meadow said.
To confirm their initial estimation, the team tested something whose input was entirely known—a computer.
Model behavior
Past studies have shown that people can expect and predict the next words in a sentence. This predictive processing is what forms the basis of large-language models like ChatGPT.
For this study, researchers built a predictive model and trained it on the data collected from the parents. They fed the model in stages, simulating how a child would hear the language.
“To test the model, we give it utterances the child produced that contained a determiner, and we block out the determiner. Then the model has to predict the word that goes in the blocked-out space,” Goldin-Meadow said. “And for the most part, it does what the kid does.”
The model also confirmed the timeframe that children start to say determiner-noun combinations that go beyond what they’ve heard: around 30 months.
“For the model, we can be very sure that it has gone beyond the input it’s gotten,” Goldin-Meadow said.
Goldin-Meadow says pinpointing moments of productivity may be crucial for understanding a long-standing theoretical question in linguistics: How much linguistic input do kids have to hear to learn particular language structures?
This is an essential question for another area of Goldin-Meadow’s research: homesigners. Homesigners are deaf children who have developed their own gestural signs to communicate. Since they haven’t had access to an established sign language like ASL, their own system of gestural language could shed light on which linguistic constructions children expect to find in the languages they are learning.
According to Goldin-Meadow, experimenting with computer modeling can test insights provided by homesigners; in this case, that homesigners are able to invent determiner-noun combinations.
“Determiner-noun constructions may be a lot easier to learn than constructions homesigners don’t invent,” Goldin-Meadow said. “And, if so, then maybe we can play around with our computational model and give it a lot less input and still have it master determiner-noun combinations.”
—Prof. Jacqueline Stewart
