The 70-Year-Old Mystery of How We Speak
Every time you string words together into a sentence, you’re improvising. You’ve probably never produced that exact sentence before, yet your mind effortlessly assembles words into a meaningful message. For decades, linguists have explained this ability by positing an internal “grammar” — a set of rules that builds sentences like a branching tree. This hierarchical model, championed by Noam Chomsky in the 1960s and built on ideas from the 1950s, has dominated the field for 70 years. It suggests we mentally organize words into constituents: noun phrases, verb phrases, and so on, nested within each other like Russian dolls.
Timeline of Key Events
- 1950s – Chomsky proposes hierarchical generative grammar.
- 1960s – Formalization of tree‑structured syntax becomes mainstream.
- Jan 21, 2026 – Nielsen & Christiansen publish their study in Nature Human Behaviour.
- May 30–31, 2026 – Major science outlets (SciTechDaily, Cornell, ScienceIllustrated) cover the study, sparking widespread discussion.
But what if we’ve been looking at language the wrong way? A new study from Cornell University and Aarhus University, published in Nature Human Behaviour on January 21, 2026, challenges that long‑held theory. The researchers, Morten H. Christiansen and Yngwie A. Nielsen, argue that language might rely less on intricate hierarchical structures and more on short, linear sequences of words — think LEGO blocks that snap together, not a meticulously designed tree. Their findings, backed by eye‑tracking experiments and real‑world conversation analysis, could reshape our understanding of how language works, how children learn to speak, and even how different human language is from animal communication.
The implications are profound. If language is indeed simpler at the mental level than we thought, the cognitive divide between humans and other species may be far smaller than previously believed. It could also change how we teach second languages and design AI systems that process natural speech. Let’s explore the evidence, the methods, and what it all means.
Tree Structures and the Grammar of Thought
Since the mid‑20th century, the dominant view in linguistics has been that language depends on hierarchical mental representations. In this framework, we don’t simply string words together in a linear “wondered if you” fashion; instead, we build sentences from the bottom up, combining smaller constituents into larger ones according to grammatical rules. Take the sentence “She ate the cake.” Under the hierarchical model, “the” and “cake” first merge into a noun phrase (NP) “the cake.” That NP then combines with the verb “ate” to form a verb phrase (VP) “ate the cake.” Finally, the subject “she” attaches to the VP to yield the complete sentence. This creates a tree‑like structure with clear parent‑child relationships across multiple levels.
| Aspect | Hierarchical Model | Non‑hierarchical Model |
|---|---|---|
| Mental Representation | Tree‑like constituents (NP, VP, etc.) | Linear word‑class sequences (chunks) |
| Example | “the cake” as NP within VP | “in the middle of the” as stored chunk |
| Processing Mechanism | Rule‑driven combination | Priming of frequent sequences |
| Historical Origin | Chomsky, 1950s–60s | Christiansen & Nielsen, 2026 |
This tree model became the cornerstone of generative grammar, with Chomsky arguing that such hierarchical organization is what makes human language unique. It explains how we can produce and understand infinitely many novel sentences: the mental grammar provides a recursive combinatorial system. The theory also predicts that speakers are sensitive to constituency boundaries — for example, they find it easier to recognize or recall phrases that form coherent constituents.
But Christiansen and Nielsen point out a puzzle: some of the most frequent word sequences in everyday speech don’t fit neatly into constituents. Phrases like “in the middle of the” or “wondered if you” are commonly used, yet they don’t correspond to traditional grammatical units. These “nonconstituents” have been largely overlooked because they don’t conform to the hierarchical expectation. “But not all sequences of words form constituents,” the authors write. “In fact, the most common three- or four-word sequences in language are often nonconstituents, such as ‘can I have a’ or ‘it was in the.’”
If language is truly hierarchical, why would these ill‑formed sequences be so prevalent? The new study suggests the answer lies in a different kind of mental representation — one that’s flatter and more linear.
Priming the Mind’s Linear Patterns
Christiansen and Nielsen set out to test whether these non‑hierarchical, linear sequences are actually stored in our mental language knowledge. Their approach combined two powerful methodologies: an eye‑tracking experiment and a corpus analysis of natural phone conversations. The eye‑tracking study measured how quickly participants processed sentences containing specific word‑class sequences. If a sequence had been encountered before, their eyes moved faster, indicating a “priming” effect — the sequence had been stored and reactivated from memory. The conversation analysis, on the other hand, examined thousands of real‑world utterances to see which linear patterns occurred most frequently.
The results were clear: both constituent and nonconstituent sequences showed priming. That is, even though “in the middle of the” doesn’t form a grammatical constituent, hearing or reading it once made people process it faster the next time. This provides compelling evidence that our mental representation of language includes these linear chunks alongside any hierarchical structures. “I think the main contribution is showing that traditional rules of grammar cannot capture all of the mental representations of language structure,” Nielsen stated.
The published paper, “Evidence for the representation of non‑hierarchical structures in language,” goes further, reporting that non‑hierarchical structures appear consistently across reaction‑time tasks, eye‑tracked reading and natural conversation. The authors conclude that language may be represented more like a collection of overlapping linear patterns than a single, monolithic tree. This doesn’t entirely discard hierarchy; it suggests that both types of representations coexist and that the linear component has been underestimated.
Importantly, the study focused on English, but the researchers believe the findings hold across languages. “Although the research focused on English, the authors suggest the results may apply to many languages and could influence future research on how language evolves, how children learn to speak, and how adults acquire new languages,” noted a summary from Cornell University.
Building Sentences with LEGO Blocks
To explain their alternative view, Christiansen and Nielsen use a vivid metaphor: imagine constructing a sentence is like building a house with LEGO bricks. The traditional hierarchical model treats language like crafting every brick from scratch and then assembling them according to a precise blueprint — a slow, rule‑driven process. The new model suggests we often use prefabricated pieces. The individual words are still the bricks, but we also have ready‑made door frames and window sets — short, high‑frequency sequences that we can insert directly without mentally parsing their internal grammar.
Phrases like “can I have a” or “it was in the” function like those prefab pieces. We don’t need to consciously apply grammatical rules to produce them; they’re stored as whole units and retrieved when needed. This explains both the speed and fluidity of everyday conversation and the prevalence of sequences that defy traditional constituency. “Under this view, speakers draw on short, linear sequences of word types, including nouns and verbs, rather than relying entirely on abstract grammatical rules,” the authors explain.
This flatter, more associative system could be computationally simpler and easier to learn. It also aligns with how artificial neural networks process language — they pick up on statistical regularities without explicit hierarchical annotations. In fact, the study’s implications for AI are intriguing: if human language uses a mix of hierarchical and non‑hierarchical representations, perhaps future language models should incorporate both.
Christiansen hinted at a broader consequence: “It might even be possible to account for how we use language in general with flatter structure. Importantly, if you don’t need the more complex machinery of hierarchical syntax, then this could mean that the gulf between human language and other animal communication systems is much smaller than previously thought.” This challenges a long‑standing belief that hierarchy is what sets human language apart.
Rewriting the Story of Language
If the study’s conclusions hold, several fields could be transformed. In linguistics, the focus may shift from purely hierarchical analyses to exploring how linear statistical learning shapes mental representations. In psychology and cognitive science, models of language processing will need to accommodate both constituent and nonconstituent memory traces. Education could also benefit: second‑language teaching often emphasizes grammatical rules; recognizing the role of formulaic sequences might lead to more natural, chunk‑based instructional methods.
The potential bridging of the human–animal communication gap is perhaps the most provocative. Hierarchical syntax has long been cited as a key distinction between human language and animal signaling systems. If many aspects of language can be captured with flatter, non‑hierarchical structures, then the uniqueness of human language may lie in the scale and combination of these patterns rather than in a qualitatively different cognitive mechanism. This doesn’t diminish human language, but it reframes the evolutionary question.
For AI, the findings reinforce the success of transformer‑based models that rely on attention mechanisms and positional encodings without explicit tree structures. “If you don’t need the more complex machinery of hierarchical syntax,” Christiansen notes, then both brains and machines might share a common computational principle: learning frequent linear patterns. Future NLP systems could become even more efficient by explicitly modeling these sequences.
Of course, questions remain. The study focused on English; cross‑linguistic verification is needed. The exact interplay between hierarchical and non‑hierarchical representations also invites further investigation. But one thing is clear: a 70‑year‑old assumption has been seriously challenged, opening new avenues for understanding the human capacity for language.
A Simpler Foundation for a Complex Ability
The evidence assembled by Christiansen and Nielsen paints a compelling picture: language may rely more on short, linear sequences than previously acknowledged. Their experiments — eye‑tracking and phone conversation analysis — demonstrate that nonconstituent chunks are not only common but also cognitively real, as shown by priming effects. This doesn’t throw grammar out the window; it suggests that mental representations include both hierarchical and non‑hierarchical elements, with the latter perhaps playing a larger role than we thought.
If validated across languages, these findings could lead to revised models of language acquisition, more effective language teaching, and a deeper appreciation of how statistical learning shapes one of humanity’s most defining traits. The gap between human language and animal communication may be narrower than the grand old tree of grammar implied, inviting us to rethink evolution and cognition.
As with any scientific advance, this is not the final word but an invitation to explore further. Future research will test the generalizability, probe the neurocognitive mechanisms, and refine our understanding of how hierarchy and linearity intertwine. For now, the takeaway is both humbling and exciting: our capacity for language may rest on simpler building blocks than we imagined.
*This article was generated by AI based on research from multiple sources. While efforts are made to ensure accuracy, readers should verify information independently.*
Post a Comment