Language is often treated as a cultural artifact: something handed down from generation to generation, shaped by history, identity, politics & social change. We talk about how languages borrow words, simplify structures or develop new expressions as societies evolve. But there is another, equally intriguing possibility. What if languages are also shaped by the minds that learn them? What if some features survive not because they are better at communicating, but because they are easier for learners to acquire, remember & pass on?
This idea has fascinated linguists for decades. Human languages, despite their immense diversity, tend to display strikingly similar properties. They are highly structured, endlessly productive & remarkably learnable. One explanation is that, over thousands of years, languages have adapted to the cognitive preferences of successive generations of learners. Just as species evolve under biological pressures, languages may evolve under pressures of learning, memory & transmission.
This idea sits at the heart of iterated learning, a framework associated with Kenny Smith, Simon Kirby & colleagues. The basic claim is simple: language changes slightly each time it passes through a new learner. Over many generations, these small changes can accumulate, creating systems that are increasingly structured, systematic & learnable.
A recent study by Devon Jarvis, Richard Klein, Benjamin Rosman & Andrew M. Saxe, published in PNAS, explores this idea in an unusual way. The researchers asked a deceptively simple question: if a “language” is repeatedly passed from one learner to the next, will structure & systematicity emerge naturally over time?
The Study
The researchers created a chain of deep linear neural networks, each representing a new generation of learners. Rather than learning directly from an original source, each network was trained on the outputs produced by its predecessor. In other words, the “language” was repeatedly transmitted from one generation to the next, creating a computational version of cultural transmission.
The researchers then analysed:
- How the networks’ output systems changed across generations
- Whether increasingly compositional structures emerged
- Whether the networks developed systematic generalisation, meaning the ability to apply known patterns to entirely new combinations
Compositionality is one of the key properties of human language. If you understand green car & red bicycle, you can usually understand green bicycle without ever having encountered it before.
Another everyday example would be knowing online lesson, grammar task & speaking activity, then immediately understanding online speaking activity. We constantly combine familiar components in novel ways.
The authors used mathematical analyses of both shallow & deep networks to examine exactly when these abilities emerge.
The findings
The results strongly supported a long-standing prediction from iterated learning research.
- Repeated cycles of learning & transmission gradually produced more structured, compositional systems.
- Networks trained through iterated learning were better able to generalise known patterns to new situations.
- Deep networks were far more successful than shallow ones, suggesting that complexity plays an important role in the emergence of language-like structure.
- Even then, systematic generalisation emerged only gradually, leading the authors to distinguish between weak & strong forms of generalisation.
In short, language-like structure did not need to be explicitly designed. It emerged gradually through repeated cycles of learning & transmission.
Why this matters for ELT
The study is highly theoretical, but it connects intriguingly with several ideas familiar to language teachers.
Usage-based approaches associated with Michael Tomasello, construction grammar, emergentist theories of language learning & research on language change all emphasise that linguistic structure can emerge from repeated use rather than from pre-existing rules alone.
The findings also resonate with observations from learner language. Learners often create productive patterns before mastering individual exceptions.
Consider how learners may initially overgeneralise-ed forms such as goed or buyed. From a traditional perspective these are simply errors. From an iterated-learning perspective, they also reflect a drive towards systematicity. Learners are searching for patterns that can be applied widely.
Rather than being a static object, language appears to be a moving target continually shaped by generations of learners.
Teacher Takeaways?
- Learners’ tendency to overgeneralise may reflect a natural search for systematic patterns rather than simple carelessness.
- Regularity often emerges before accuracy. Building productive patterns may be a necessary step in acquisition.
- Language systems that survive over time may do so partly because they are learnable, not simply because they are communicatively efficient.
One of the most fascinating implications is that language may be shaped as much by the minds that learn it as by the people who use it.
As English is learned by more people than ever before, what changes might future generations leave behind?
Leave a Reply