What goes on in our minds when we construct an utterance? Linguists often use syntax trees to represent the structure of sentences, but are they psychologically real? Yngwie Nielsen and Dr Morten Christiansen have found evidence for something else: we can recognise patterns in strings of words, even when they don’t form coherent “treelets”. They’re giving us a walkthrough of their latest work.
Timestamps
00:00 Start
00:31 Introductions: Yngwie and Morten
05:19 Insights into linguistics communication
07:45 What are syntax trees?
09:13 Why linguists love syntax trees
14:15 Treelets vs chunks: Looking beyond hierarchical structure
17:46 Wanna and gonna: Words that cross treelet boundaries
22:43 How to prime someone
28:18 Priming in this experiment: People do recognise chunks
32:26 Are people just filling in the treelet blanks?
35:23 Were they accidentally smuggling in treelets?
38:47 Do we process both treelets and chunks?
42:23 DensiTrees: A way of representing fuzzy networks
44:01 What are we doing mentally when we make an utterance?
47:20 What is language for?
49:29 Grammatical glue: How do we connect chunks?
53:23 Being able to language is bonkers
56:30 Should we be studying language differently?
01:01:09 Wrap-up and goodbyes
Listen to this episode
Video
@becauselangpod New ep: What's going on in our minds when we make sentences? Are we constructing trees? Or do we think of words sequentially? We're climbing down from the trees with Morten Christiansen and Yngwie Nielsen. Video (recommended): https://www.patreon.com/posts/video-137-are-157061800 Audio: https://www.patreon.com/posts/audio-137-are-157053696
Patreon supporters
Thanks to all our patrons!
Become a Patreon supporter yourself and get access to bonus episodes and more!
Become a Patron!Show notes
Nielsen, Y. A., & Christiansen, M. H. (2026). Evidence for the representation of non-hierarchical structures in language. Nature human behaviour, 10(3), 579–588.
https://doi.org/10.1038/s41562-025-02387-z
https://www.nature.com/articles/s41562-025-02387-z
Research briefing: Priming of non-constituents reveals linguistic structure beyond grammar
https://www.nature.com/articles/s41562-025-02388-y.epdf
Densitree
https://www.cs.auckland.ac.nz/~remco/DensiTree/
McCauley, S. M., & Christiansen, M. H. (2017). Computational Investigations of Multiword Chunks in Language Learning. Topics in cognitive science, 9(3), 637–652.
https://doi.org/10.1111/tops.12258
https://pubmed.ncbi.nlm.nih.gov/28481476/
Fedorenko, E., Piantadosi, S.T. & Gibson, E.A.F. Language is primarily a tool for communication rather than thought. Nature 630, 575–586 (2024).
https://doi.org/10.1038/s41586-024-07522-w
https://www.nature.com/articles/s41586-024-07522-w
Transcript
[Transcript provided by SpeechDocs Podcast Transcription]
MORTEN CHRISTIANSEN: Daniel, I have to ask you this. It looks like you’re… I’ve noticed in these, it looks like you’re outside, but you’re not.
DANIEL: [LAUGHS] People often say that.
MORTEN: I take it because it’s dark. It should be dark where you’re at now.
DANIEL: It is dark where I’m at. So, here’s what it really looks like. [REMOVES BLUR] It looks like a pantry. [LAUGHTER] Check it out. This is where the magic happens.
[BECAUSE LANGUAGE THEME]
Hello and welcome to Because Language, a show about linguistics, the science of language. I’m Daniel Midgley and I’m here with Hedvig Skirgård. Hi, Hedvig.
HEDVIG: Hi.
DANIEL: This is going to be one of our shorter standalone episodes. We like to do these when we have a topic that we really like to dig into. And sometimes, it’s because it’s a topic that’s more current and breaking. Sometimes, it’s because it’s more theoretically challenging, and I feel like this one is both. We found some research that we think is really exciting because of its implications for how we study language and how language works, but we need needed some help. And so, we have contacted the authors of this work. We’d like to welcome Yngwie Nielsen of Aarhus University. Hey, Yngwie.
YNGWIE NIELSEN: Hello.
DANIEL: And Dr Morten Christiansen of Cornell University. Hi, Morten.
MORTEN: Hi. Hi. I should mention I’m also at Aarhus University, by the way, part time.
DANIEL: I read the affiliations on your paper and both of you are in three different places.
MORTEN: That’s true.
HEDVIG: How does that work? Do you have like a quantum state, or…?
MORTEN: [LAUGHTER] Yeah, I do sort of switch back and forth in between. Rather, I’m there a few weeks a year, and then obviously we work remotely as well.
DANIEL: Okay. But now, you’re being observed. So, the waveform has collapsed and you’re both here with us. That’s really cool, okay.
Morten, this is your third show with us, which means you’re an honorary cohost and you get to crash any episode you want. Do you want to tell us what kind of linguistics or what kind of language you’re really into in your research, or just for fun?
MORTEN: Well, right now, I’m actually working on a book on conversation. So, I’m really into how language actually works in real life, in conversation, how are we actually using it in its sort of original medium, as it were, in face-to-face interaction. Sort of almost like what we’re doing right now. Of course, we’re not actually physically face-to-face, but actually visually still we can see each other and so on. So, this is the kind of use of language that I’m really interested in. And it also bears on some of the things that we’ll be talking about today, I think.
DANIEL: Now, are you into this kind of thing because it gets us out of the abstraction and gets us into like brain stuff, like how we make decisions and how we perceive? Is that what does it for you?
MORTEN: No, not really in that direction. I’m more interested in understanding language as it originates in conversation. That is, what is it about our ability to use language in the here and now, in talking to each other like we’re doing right now in real time? What does it mean for how we should think about the nature of language, the mental representation of language and the cognitive underpinnings of language? That’s kind of what the book is all about. So, it’s kind of rethinking language from the viewpoint of conversation.
HEDVIG: That’s a very usage-based perspective, a very functional perspective. So like, what language is doing when we are using it in the medium that we’re using it in, have been using it for the longest amount of time. Text and these abstractions are relatively new things.
MORTEN: Yeah, that is… Yeah, exactly.
DANIEL: Okay. Yngwie, first time guest, welcome to the show. Tell us about your work and what’s it about.
YNGWIE: Thanks. Well, that’s a great question. It’s something I’m currently pondering myself being in the middle of the PhD. Certainly, [LAUGHS] part of it is on the structure of language and this sort of remarkable ability that humans have to put together words in ways that are sometimes creative, sometimes new, but also often based largely on their own experience with the language. And just taking a basic psychological perspective on that. It’s a very remarkable ability. Of course, it’s something that we get to practice a lot more than, say, riding a bike or whatever else kind of skill you have. It’s something that you practice from very early on and that you get and throughout your entire life. So certainly, it’s not surprising that we’re able to do exciting and complex things. And accordingly, the knowledge that you must obtain is also of a very complex and very interesting nature. So, that’s what I’m sort of trying to probe, I guess, is what is this knowledge exactly?
HEDVIG: And I think the fact that everyone uses language all the time is both a blessing and a curse when you’re working on languages, because it also means that various members of the general public also have a lot of opinions about how they think linguistics and languages work, whereas people don’t have as many opinions about how jet engines work [LAUGHS] because it’s not something you do very often. At least that’s what I found. So, it’s nice because you can often connect with people because you can say, “Oh, have you thought about this?”, because basically everyone has, or, “I use this language,” but it can be a blessing and a curse.
DANIEL: Now, this paper has been published in Nature Human Behaviour. It’s called Evidence for the Representation of Non-Hierarchical Structures in Language. Now, if that’s going to make anybody in the audience bail, please don’t, because I think this is important stuff and we’re going to make it easy and fun. But first, Morten and Yngwie, is it difficult for you to explain your work to a nonlinguistic audience? Have you tried, and what secrets can you impart?
YNGWIE: Oh, do you want to start, Morten?
MORTEN: Well, it is tricky and this sort of paper, despite its importance, it’s a bit tricky. We did try writing a sort of a short, more accessible version of the paper that was also published as a research briefing in Nature Human Behaviour. But yes, it is tricky on the one hand, so explain why do people like trees, for example. What does it mean for what we actually do with language and so on, but I think we can sort of unpack that, all four of us together here today, if we give it a try.
HEDVIG: Yeah, exactly, because that’s the first thing you’ve got to explain to a general audience is that linguists like to draw trees of words in a sentence, which doesn’t come naturally to everyone. I’ve taught Intro to Linguistics, and it takes a while for people to get the idea of hierarchical structure and then to present them with the idea that actually maybe we should critique that view. I think if you pitch it as like a lot of linguists think this and actually it could be this. That part can be intriguing in itself, I think.
DANIEL: Okay, so we’re starting to talk about trees. But before we get there, Yngwie, your insights so far into linguistics communication?
YNGWIE: Right. I think I’m also looking for the secret of how to make my work interesting, especially to my parents. I think a lot of people, when they hear that it’s grammar adjacent, the connotations that that bears is the first challenge. And then, the second challenge is that we all have so much experience with language, and I think we take a lot for granted about how it’s done and possibly also about how complicated it is once you start digging into it a little bit. So, I think at least that’s the approach that I’ve been trying to take, is unpacking the complexity first and then moving on to why we believe the things we do.
DANIEL: Okay, well, the tree structure idea is certainly one of the more complex things that I’ve tried to get to my students. So, maybe let’s talk about that for a second. There are lots of different ways you can tree a sentence, but I guess one of the most common is you start with a letter S at the top, and that means sentence. And then, it breaks down into two nodes. The first one could be an NP, noun phrase, and the second one could be a VP, verb phrase. And then, you break it down and break it down. And then at the end, the leaves are words. But by showing that some groups of words are part of a subtree or maybe a treelet — I’ll call them treelets — You can see the little treelets in the tree.
HEDVIG: I was going to say exactly… that I tend to do it the opposite way. So, start with the words in a sentence like, “This is a beautiful dog.” And then, say, “Okay, are there any words in the sentence that you think are more closely related to each other than to other words in the sentence?” And then, sort of do a grouping and then say, “Ah, this grouping is…” You can understand it as hierarchical and then we can draw a tree so that BROWN is closer to DOG than it is to THIS. That’s usually the way I go about it. Most Linguistics 101 classes introduce you to the basic ideas of trees. And then linguists diversify into lots of different ways of drawing those trees and movement in those trees and abstract theories.
DANIEL: I want to ask why we like trees so much. Why do we love them? I love them. We use them a lot.
MORTEN: Yeah, I mean, who doesn’t like trees? I mean, we are depending on trees for part of our oxygen along with other organisms. But I think one of the things that’s interesting is that the focus on trees that there is in contemporary linguistics, it’s actually a relatively new one. So, the first person, as far as I can work out, to talk about trees that have like sentence node, you start with that, you describe a sentence as a tree-like structure, to sort of try kind of characterise the hierarchical structure of it, that started with Wilhelm Wundt. He put that forward in at the beginning of the 20th century as a kind of way of describing the mental structure of language. Now, that was picked up by the structuralists, of course, and then subsequently it fell a bit out of favour with behaviourism, but then it came back very strongly with Chomsky and others afterwards.
But prior to that, people didn’t really think of sentences or language in terms of these hierarchical structures. There were some earlier sort of very sort of proto-hierarchical structures in the sort of the Middle Ages and the Renaissance perhaps. But in general, it’s actually a relatively new thing and so I think that’s important to keep in mind.
And of course, there’s other areas where we think of trees too. So, for example, oftentimes, we think of, say, evolution in terms of trees. You have these evolutionary trees, but it turns out we know now that maybe that’s not actually the best way of thinking about. And then, actually, sort of as a little side note, one of the things that’s really interesting is that Darwin kind of borrowed the notion of descent from linguistics from Schleicher.
DANIEL: What?
HEDVIG: Schleicher?
MORTEN: So, Schleicher in 1853 was the first person to indicate sort of relationship between languages. So, this is not within a language, but between languages, how some languages are more closely related to others using a tree. And Darwin actually sort of was aware of that going on in linguistics and even referred to that in his notes and in some of his writings. So, when three years later, in 1859, in Origin of Species, he presented the notion of a tree as a way of indicating how a species might be related to one another, evolutionarily speaking, he was borrowing from linguistics at that time.
Of course, now we know that actually that way of thinking about trees and evolution is actually not completely the right way because there’s a lot of what’s called horizontal gene transmission going on, so between species that are not related. So of course, that happens a lot in bacteria, but it happens even in our ancestry as well. That’s why, given that we are all of Northern European descent, all of us have a certain amount of Neanderthal genes in our genome. That actually comes from crossbreeding across species, which, according to these strict trees, shouldn’t happen. So really, what you get instead is these reticulated trees.
And it turns out that I think also when it comes to actual language and the sentence, we might be thinking about the representation as a network rather than as these beautiful trees. So, I think the reason why we like trees is it can actually be quite useful in terms of describing the structure and getting people to understand that, yes, there are words that go together in certain ways in the way that Hedvig was talking about earlier, that they can be quite useful at a descriptive level. But it’s a completely different matter to say that this is actually what’s going on in the head as such.
So, I think we shouldn’t throw the trees away. No, we’re not suggesting that. And they could be quite useful in getting people to understand there are some interesting relationships between words and sentences and so on. But I think it’s important that we might want to rethink whether tree-like structures is also what’s in the head when we’re using language, like what we’re doing right now.
HEDVIG: Yeah, I think that’s really important to know why you’re applying a technique. I have met people who like to draw trees or sentences who say, “Oh, I think this is an efficient computational way of explaining the structure I’m after.” But when asked, “Oh, do you think humans do this in their heads?”, they say, “Oh, I have no idea. I make no claims on that.” And for me, as a more like functional usage-based person, I’m kind of like, “Well, in that case, I don’t know if I care that much.” [DANIEL LAUGHS] In that case, we’re just like solving puzzles and maybe teaching computers how to emulate us or something like that, but we’re not really explaining what humans are actually doing.
As a small side note, at the institute I work at, Max Planck Institute for Evolutionary Anthropology, there’s a lot of research on Neanderthals, and I recently was sampled for my DNA and I’m going to find out how much Neanderthal I am. I’m very excited.
DANIEL: Oh, we’ll look forward to that.
HEDVIG: Yeah. I’ll tell you my test result.
DANIEL: So, if we’re not looking at trees, if we’re moving away from trees or treelets, I’ve always thought that we represent sentences as trees, but we hear the words sequentially, almost like they’re beads on a string. They seem to come in chunks. So, one thing I’ve gotten from the paper is we’re kind of contrasting trees in which the words are constituents versus chunks, like pieces of text that are just simply next to each other without any claims on that hierarchical structure, like beads on a string. Am I getting that right from the paper?
YNGWIE: Well, it’s a good question. So, it’s hard to get out of the tree-like thinking. You need some kind of other frame of mind. And indeed, chunks is possibly one way. It’s another way of saying linear or sequential structure or serial structure. So, yes, things that are next to each other, though I think by just saying next to each other, we might also be capturing some language phenomena that might not be in the head. So, it’s a little difficult to say exactly where we put this “chunk” term. Certainly, if you look through a corpus and you look through a text, you can find lots of things that are next to each other that are perhaps not necessarily part of the knowledge we have of a language, though certainly many of them are.
DANIEL: Okay, so just to make it super clear… Oh, go ahead, Hedvig.
HEDVIG: No, I was going to ask if we can maybe take an example sentence.
DANIEL: I have a sample. So, for example, “Here is the red book that my father bought.” So, that’s a sentence and we hear the words one by one, but we could represent it by a tree. But your paper mentions constituents. Like, for example, in the sentence, “Here is the red book that my father bought,” “the red book” would be a constituent because that would fit as a treelet and there’s no bits of that three-word phrase that go outside of a treelet as far as I would draw it. Or “my father bought,” I think that would be a treelet as well. I think that would be a constituent because… it’s a treelet, depending on how you draw the tree. But “book that my,” those are three words that I think would cut across different treelets, so they’re not a constituent. They’re just three words.
So, in the experiment, as I understand it, you might want to give people these three words to see if their brains go, “I have no idea about the internals of this,” or, “Yep, I can totally see those three words as a unit,” even though they’re not a treelet. Am I kind of getting there?
YNGWIE: Yeah. So that’s certainly half. That’s part of it is you can see that people do treat these chunks like units, at least to the same level of evidence that we think people treat words like units which is again an assumption. So, for instance, if you look at how frequent the phrase is, well, that predicts how good you are at processing the phrase over and above how frequent the component words and component phrases are.
Moreover, how meaningful it is. Now, it might sound a little strange to talk of how meaningful things that “book that my” is, [LAUGHS] but there is indeed a little variation if you ask people in how meaningful they are and that too matters in terms of how people process them. So, there does seem to be a sense by which, yes, they are treated as units, at least to the same extent that we can claim that words are treated as units.
HEDVIG: Can I do a test on the sentence, “I want to dance”? So, a lot of three structure that people would propose would put “to dance” as one unit and “want to dance” as another unit, and then “I want to dance” as the whole unit. I think that’s how most structural linguistics classes would group those, TO belongs more to DANCE than to anything else.
However, if we look historically, we have actually merged WANT and TO. You could argue WANNA, and that arguably crosses between two different constituents, two different tree… Did you say treelets?
DANIEL: I’m using the term “treelets.” I hope that’s okay.
HEDVIG: Yeah. So that’s an example maybe of something that is not only super frequent so that people treat WANT TO as one unit, but they’re so frequent even that they make it into a word. Is that the kind of things we’re looking for?
MORTEN: I think that’s one of the effects of when you have these sort of frequent occurring sort of chunks or subchunks that they can sort of merge together. So, there’s a whole field of linguistics called grammaticalisation that’s all about how these kind of phenomena can show up but that’s probably a story for a different podcast at some point.
DANIEL: Love grammaticalisation.
MORTEN: But part of that… I mean, part of the theory is that part of that comes from repeated use that it becomes compressed both in terms of the duration, but also it takes on its own role. So, also very interestingly, you can use it in certain contexts, but not other contexts. So, for example, there’s another example like going to that GOES into GONNA. So, say, “I’m gonna go to the store,” for example. But what you can’t say, “I’m gonna the store.” So, you can say, “I’m going to the store,” but you can’t say, “I’m gonna the store.” But they take on some new roles as well.
So, language is wonderful in the way it changes all the time. And grammaticalisation is one of the ways in which you get these changes that come into play. And part of that is driven by, at least according to the sort of the people who do work on this, is sort of on the frequency of these items as they occur and so on and they kind of get shortened and can take on different meanings as well also.
DANIEL: And GONNA is really tightly fused. Like, when I tell my daughters to stop watching TV, “Will you please stop watching TV?”, they can’t say, “I’m going.” They have to say, “I’m gunnu. I’m gunnu.” It turns into something like gunnu.
HEDVIG: Gunnu. That’s weird.
DANIEL: I was gunnu.
HEDVIG: So, GONNA and WANNA or actually WANT TO and GOING TO are examples of things that straddle different constituents. So, they’re the kind of candidates that your paper is addressing. Can you tell us more about the experiments and how people behaved?
YNGWIE: Sure. So, what we talked about before, I guess, is it treated as a unit, right? Is the chunk treated as a unit? That’s only half of the story. Because we sort of wanted to say, “All right, we’re building on work that says it’s treated as a unit, but are we also generalising some sort of structural knowledge across different chunks.” So, in that case, it’s not enough to say, “Here we can see it being treated as a unit.” We need to show some sort of generalisation, if that makes sense.
DANIEL: Okay, so what we want to know then is, do people perceive these chunks that straddle treelets, are they able to perceive them as part of a pattern?
YNGWIE: Exactly.
DANIEL: So, let’s give them beads on a string and see if they can recognise them as valid things. And if they can, then there’s no need for us to presume that we process language in tree-like units. It starts to look like beads on a string are actually real.
YNGWIE: Right, exactly. Or to borrow the phrase from before, as real as you can show the trees to be. So, we can’t claim to show that there are no trees, but we can claim to show that there is more than just trees and then the rest becomes how do you… What do you make of that?
DANIEL: Okay, so we’re not trying to overturn… We’re not trying to chop down all the trees. [LAUGHS] We’re trying to show that there’s a little bit more.
MORTEN: Well, we are, but we’re not there yet. [LAUGHTER]
HEDVIG: Oh, we’ve tapped into their secret agenda, Daniel.
MORTEN: But also, still, trees can be useful on the descriptive level. But what we want to do is get rid of the trees in theories of how we process and use language, but we’re not there yet, and it’s science. We’re trying to do science. So, we can only say as much as our evidence allows us to. This is what Yngwie was pointing out.
YNGWIE: Exactly.
DANIEL: Okay.
HEDVIG: Very humble, very good.
DANIEL: Well, from reading through the paper, these experiments revolve around a lot of priming. So, can I tell you what I know about priming and then you can tell me how you used this. My understanding of priming is that there have been experiments where you get somebody in the lab and you say, “Right, we’re going to show you three things, but don’t worry about the first two things. Only worry about the third thing. We’re only focusing on the third thing. And when you see the third thing, hit a button as soon as you can, whether it’s a word or not a word. Word or not word. That’s all you got to do for the third thing. Just hit a button as soon as you can, whether that thing is a word or a nonword.”
And then, you show them three things. So, you go • not a word, • not a word, • word. And they go, “Oh, that’s a word,” and they hit the word button. But maybe somebody else will get it like this, • not a word, • word, • word. And they’ll go bop, the third thing was a word. But they’ll do it slightly faster because that second thing was a word, and that sort of paved the way for them to say that third thing was a word.
YNGWIE: Yeah, yeah. So, priming is a word that’s used by many people in different ways. But what you’re pinpointing there is exactly the way we think of it. You can think of it as sort of very, very recent exposure or very, very recent experience somehow facilitating what you are doing in the moment.
HEDVIG: If you’re friends with a linguist, you might have experienced that they want to know how you say a certain word, but they don’t want to say the word because they don’t want to prime you to pronounce it a certain way. So, instead, you might say, like, “Oh, the fruit that grows on a tree that’s mostly red, and you put in a pig’s mouth,” and you want them to say, “Apple,” but you want to know if they say /æpəl/ or… I actually don’t know any variations of pronouncing APPLE, I realised. And if you have linguist friends, they might do this to you, and you feel like it’s like a riddle, but they’re trying to not prime you because they don’t want you to repeat what they said.
YNGWIE: Exactly.
MORTEN: So, priming is something also that people have studied in psychology for a long time. And part of the motivation for looking at priming is the idea that if you can prime something, then we have some sort of representation of whatever it is you’re priming in the head.
So, for example, one of the things that have been shown that if you come across a particular word, say, CAT, and then there are some other words that come across, and then you come across, not long after that, a CAT again, what happens is that you’re faster at recognising and processing CAT the second time around, because the assumption is that you have some sort of mental representation in your head of CAT, and that is the one that it’s still sort of, perhaps partly sort of active in your head, and so it’s easier to access the second time around. And that’s the kind of priming that we took advantage of in our experiment we can tell you more about.
DANIEL: Yeah, please. How did you do it? And let’s start with the first study. How did you use priming here? It wasn’t word versus not word.
YNGWIE: Exactly. So, we needed to work with chunks instead. So, we borrowed a task from the chunk literature that’s actually very similar to the word/nonword example you gave before, where instead of deciding whether a word or a string of letters is real or not, it’s a string of words and you have to decide not whether they are real or not, but whether they could be used in a sentence. So, I’ll give an example, for instance, “book that my,” can be used in a sentence, whereas “Am that the,” can’t be used in a sentence. At least most would judge.
DANIEL: It would be a very weird sentence.
YNGWIE: It would be a very weird sentence. So now, we have a paradigm suddenly where people are making decisions on these chunks. And importantly, it’s not just nice chunks, it’s also the slightly annoying chunks like “book that my.” But that parallels the style of priming that you mentioned before. So, that’s part of it. That’s one part.
But the other part, which is the important one is you need the generalisation. So normally, the way that’s done is you take two sentences that share a structure, that share whatever it is you… What type of generalisation that you want to probe and then you try to make sure that they don’t share anything else, which is a very difficult task. So, in our case, it would be two chunks that sort of in some sense share the same structure, but that don’t share any words.
HEDVIG: So, “book that my” and “husband who your”.
DANIEL: Yeah.
MORTEN: So, these examples are great actually because they share the same what’s called parts of speech, meaning parts of speech is something like verbs or nouns or pronouns and so on like in this case here and sort of possessive markers depending on how you want to describe it. And so, these are classes of words rather than individual words and that’s actually what was used in the experiment.
DANIEL: Okay, so you gave people a really gnarly impossible sequence of three words, can’t be used in anything. That was number one. Number two, they got something like “example of an”. That’s a different sort of thing entirely. That’s a different pattern. And then, third, they got “wondered if you”. Verb, conjunction, pronoun. And they had to say is that possible? And they said yes.
And the people that got “example of an”, they knew that “wondered if you” was a possible thing. But if they had instead gotten as part two, “knew that she” — now that’s a chunk that has the same structure as the target one, the number three one — and when they got a number two that was the same kind of thing as the number three, they were like: boom, a little bit faster, “I know that’s possible.” Have I got that right?
YNGWIE: That is exactly right.
MORTEN: The only slight difference is that in the control condition, we wouldn’t start with a nonsense sequence because you can have…
DANIEL: Oh, okay.
MORTEN: So, it would always be a real, but it would be a very different one. They would have a different underlying order of these word classes. So, it could be like the “boy gave the”, for example.
DANIEL: Okay, so in other words, some people were like pattern A, pattern B, pattern C. Yep, I can tell pattern C is possible. But then, people who got pattern A, pattern C, pattern C, boom. They knew it was possible, but faster than the last group. And so, I guess that tells us that they were able to generalise across patterns even though those chunks weren’t treelets, they were something else.
YNGWIE: Yeah, so that’s the claim, like that you see some sort of generalisation here. And the question then becomes, what is underlying this generalisation? What kind of knowledge, what kind of regularity have these language users picked up on that allows them to reuse a little bit of their processing on the second trial when working on the third trial? And that’s a great… That’s the question that we are working on. But what’s important here is, as you say, it’s not the treelet knowledge that comes into play here, because these are not treelets.
HEDVIG: Mm-hmm. What were the… How do you say, the red herrings? Like, the sequence of three words that was not a unit at all that they were exposed to, what did they look like?
YNGWIE: Yeah, so there would be three words randomly put together. And most of the time when you put three random words together, it doesn’t work out.
DANIEL: Wanted of to.
HEDVIG: Wanted of to.
YNGWIE: But of course, every now and then, it does work out. So, we had to… Myself and some others, we went through and checked that they were indeed nonsense. But actually, for a regular person, of course it’s quite difficult to think of a sequence of words that is complete nonsense and cannot be used because that’s just a task that you never have. You never really have to think of something that’s completely nonsensical and unusable, but computers are great. [LAUGHTER]
HEDVIG: It’s kind of hard, yeah. And did you also have a test condition with what we’re calling treelets?
YNGWIE: We did indeed run an experiment where every now and then it was a treelet and every now and then it wouldn’t go here with the treelet. And what we find here is we actually find just very comparable effects. So, yes, they are faster. Yeah, there’s not much more to add there. We weren’t expecting the treelets not to work. [LAUGHTER]
HEDVIG: Yeah, no, of course.
DANIEL: Okay, because they’re possible.
YNGWIE: Right.
DANIEL: Very possible.
YNGWIE: Yeah.
DANIEL: Okay. Now, one possible objection, which I know about because I read this and you tested for it, was it only worked because our brains hear the three words and then sort of anticipate a fourth thing. And that fourth thing that we’re anticipating is part of a treelet.
HEDVIG: So, “wondered if she… liked me,” if I was like completing it in my head, then I actually make a full constituent, make a treelet. And maybe that’s what I’m actually operating on. Interesting.
DANIEL: Maybe we’re imagining treelets. Yeah.
HEDVIG: Wow. You’re being real role models for intellectual humility by coming up with critiques of your own work in your own work, that’s very nice…
DANIEL: You’ve got to test everything.
HEDVIG: …cool. And also, emotionally mature is what I’ll call it.
DANIEL: Okay, so how did you overcome that?
YNGWIE: We tried two things. So first, we tried to develop a slightly more sophisticated experiment that could work with this. But then later, we actually also tried to analyse sort of slightly more real-life data where we looked at full sentences and where we could… So, because we would have the full sentence, we would be able to determine exactly what constituent these sequences would appear in or not appear in.
But for the experiment, we did the following. We found four-word sequences that were constituents, but in which the first three words sort of didn’t cohere with the treelet structure. So, that would be, for instance, “live in a car”. Now, “live in a car” forms a nice treelet structure, whereas “live in a” does not. And essentially, we just redid the experiment, but every single sequence was like this, and it was random whether you saw one or the other.
So, sometimes you would see the full version, sort of what we might expect people to represent if they’re able to complete and sort of predict how would this full treelet look like. And sometimes, they would just see the little fragment. And it turns out curiously, that even if you show people the full, in this case, four-word treelet, it’s not better than if you just show them the three-word non-treelet.
Now, if it was based on sort of trying to figure out how does this fit into a treelet structure, then you would expect them to be slightly better if they were actually given the solution, so to say if they were given the full four-word treelet.
HEDVIG: Yeah, that’s interesting.
DANIEL: But they weren’t.
YNGWIE: But they weren’t.
HEDVIG: But they weren’t. And how many people did you experiment on? And you ran this experiment in English, correct?
YNGWIE: Well, we ran four experiments. The first two had 40 participants each, all undergraduates, and then the second two had 200 participants each, sourced from an online platform.
DANIEL: Now, is it possible that you accidentally smuggled in treelets? Like the groups of three words that you were presenting to them secretly were constituents in some way that you didn’t realise?
YNGWIE: That is a great question. And it turns out that at least in the first experiment, we did accidentally sneak in some. For instance, we snuck in “of the best”.
HEDVIG: Oh.
DANIEL: That sounds like a treelet.
YNGWIE: It sounds like a treelet, exactly, because it can occur as, “He is one of the best.” It can also occur as a non-treelet in “He is one of the best dancers.” So, that was actually the big challenge was that we had people work on these three-word sequences that appeared out of context. But whether or not a sequence of words is a treelet depends strongly on the sentence that it occurs in.
So, once we became aware of this, what we had to do was essentially find sequences that almost never or never occur as [CHUCKLES] treelets. So, we had to really just look through a very large corpus and determine whether or not these sequences could occur. And there are some that never occur as treelets. That would be, for instance, sequences ending in “in the.”
DANIEL: Yeah, that makes sense.
MORTEN: Unless you’re talking about the band The The.
DANIEL: Which causes a lot of havoc.
HEDVIG: Well, then you’d expect two of them.
MORTEN: Yeah, but that would be an extreme exception, I suppose.
HEDVIG: Band names do seem to be… Band names and like book titles, I don’t know, I was always weirded out by ‘Of Mice and Men’ because it’s a part of a longer phrase that I didn’t know. And I was like, “Why is there a book that’s called ‘Of Mice and Men’ that doesn’t make any sense?” And you always hear about like punk bands or something that are just very strange combination of words. It’s very poetic. But setting those aside, I assume you looked at some corpora of some more normie sentences.
MORTEN: Yeah. We did.
DANIEL: Okay, so it looks like you’ve dealt with a lot of objections, and you found that when people see chunks, it makes recognising chunks easier, indicating that chunks are real.
YNGWIE: Yeah. And the structure that they embody is a real tool. Yeah.
MORTEN: I think that’s an important point that Yngwie was mentioning here, is that it’s not just chunk of individual words that we sort of were working with in this paper, but it’s actually sequences of these word classes rather than individual words. And that’s important because that adds an important level of abstraction to these generalisations that hasn’t been shown before and that we’re able to show both for what would conform to treelets, but also what would not conform to these treelets. And that’s what we think is sort of an important contribution of this paper, that we’re not suggesting that you only are doing these sort of beats on a string with individual words, but actually seems like those beats would be more like word classes in this case if we’re talking about that.
DANIEL: We had a chat recently with Dan Parker of the Ohio State University on a recent episode. And one idea that came up is that we might kind of be doing both or multiple ways of processing. Sometimes, we might do a kind of shallow processing, which seems a bit more like chunks. But sometimes, we need to do some deep processing, which we would handle a bit more like trees or maybe treelets and what we would like to know is the conditions under which we do each kind of processing. So, is there space for both kinds of representations? Can treelets or trees and chunks coexist here?
MORTEN: Here, I would say that when it comes to the actual mental representations, I don’t think that treelets exist in our heads. That’s my view.
DANIEL: Spicy!
MORTEN: Now, I think we can, if we’re doing sort of more deliberate work and so on, we can analyse it in terms of these trees and so on. It can be a helpful way, as we already talked about, to sort of understand sort of the patterns of language and so on.
But I think when we’re using language in the here and now, like what we are doing now, I don’t think that they play much of a role. And I think most of the times, when people talk about more subtle use of language, these are oftentimes very contrived experiments. They’re not about using language in the context of a conversation, for example, because the pressures on conversations are actually incredibly tough on us because you have to respond within just a few hundred milliseconds of when somebody else finishes their turns. And so, that means you actually have to start preparing what you want to say before the other person has finished what they’re saying and that requires a lot of work, especially given that we have these severe limitations on our memory. And that means that we have to rely on this sort of more chunk based, more shallow ways of doing things.
And certainly comprehension. Most of the time you can understand pretty well what somebody else is saying, even if you miss one word or two words and so on and even if you don’t go into sort of a deep analysis of what they’re saying. And it’s only when we as psycholinguist or linguists, we create these very complex sentences that have all sort of weird stuff going on in them that people can, if they have enough time, they might be able to figure it out, but most of the time they actually don’t really understand what it is that we’re presenting them with.
And so, at least for me, and this is my hypothesis — and again I want to stress it’s a hypothesis — that we don’t actually use those trees. And I think one of the key things I think is important also with this paper is that the idea that we have these kind of tree-like structures in our head, that’s a hypothesis. It’s not a fact, although sometimes it’s actually treated like it is a fact. It’s assumed beforehand. But it is a hypothesis about the mental representation of language, that is how it is that our brains allow us to use language the way that we do.
It’s a hypothesis because we can’t really read trees of the brain as such. We can’t take a scanner or anything like that and find them. So, we have to try to deduce them from the evidence that we can get from behaviour, from looking at brainwave patterns and so on, but it’s a hypothesis.
And my hypothesis is that actually we don’t need it. So, what I’m trying to do in my work is how far can we get if we don’t assume that? Now, I could be wrong, I’m perfectly happy to concede that, but so could the people believing in in the treelets.
HEDVIG: I found it was really interesting what you said earlier, Morten, about links between different kinds of tree metaphors across different disciplines. And you talked about historical linguistics. And in historical linguistics, there’s a more and more common emerging view that if you take a bunch of languages and a bunch of words and you study which ones are common across the languages, you can make a tree, but you can also make like a more fuzzy sort of network.
If anyone is listening to this show and you google the word, DensiTree, D-E-N-S-I-tree, you can see these kinds of illustrations where you can see lines between the different languages. And sometimes all the lines are overlapping, meaning that the different analyses of trees are converging on one tree, but sometimes it’s fuzzy and we usually interpret that as something like horizontal transmission, a sort of nontree-like pattern.
And if we took the words in a sentence and drew lines between them for their relationships, it might be that quite often they end up in these sort of constituent chunks, like the brown fox. But every now and then, you get a link between WANT and TO and I don’t know what these links would be or maybe something like what would it be co-occurrence or like prediction, like how people would predict the next word, something like that.
I don’t know exactly what they should be, but maybe a lot of the times you can simplify it into a tree structure. But we know in historical linguistics that trees are simplifications and there is actually other relationships between words and other relationships between languages that we maybe are now at a state of the field that we can sort of consider as explanations or worthy of study as well.
DANIEL: Okay then. In that case, let’s just pretend that I want to say a thing. If I’m not building trees in my head all the time — and I don’t think I am — what am I doing exactly? Like, let’s say that I want to ask you for a peanut butter sandwich. And in the generative grammar approach, I would probably say, okay, let me choose my verb first, WANT, because that’s going to have a lot of the structure in it. It’s got arguments, and they’re going to be I and peanut butter sandwich. And I assemble them in the right places, make a tree, and then I say the sentence, “I want a peanut butter sandwich.” But if I’m not doing that, what am I doing when I put together a sentence? Is it constructions all the way down?
MORTEN: Well, I am a big fan of constructions, so I do think constructions are there. But I think there’s more than constructions. That’s at least what we’re also suggesting with the current work. Because I think one of the shortcomings of much work in construction grammar is it doesn’t really have good way mechanistically to put the constructions together.
And I think part of the suggestion is that these sort of non-constituent sequences of these word classes might provide sort of processing glue between them. So, what we’re doing is that when we are trying to say something, we sort of kind of grab at what are the most useful chunks that we have available to us to say what we want to say. So, rather than, for example, putting together “the red book,” we might already have an existing chunk, so we can use that, and that makes it much easier to do.
And actually, a few years ago, with a former graduate student of mine, Stewart McCauley, we created what we call a chunk-based learner. This was a computational model that actually discovered chunks from being fed child-directed speech, and then it was able to do language production by using these chunks in order to sequence the output in a way that would be consistent with what a child would say, and so on.
Of course, there’s much more going on that you have to have some idea of what you want to say. You use semantics. You also use the context in which you’re making that statement. So, you might say… if you’re in a scientific context, you might talk about felines rather than cats. But if you’re just talking with your family, you will surely say CAT rather than FELINE, because FELINE would be odd to say referring to your cat, for example. And so, you use context to figure out what are the right kind of chunks or words that you want to use at any given point.
And of course, there is structure to it, and that’s an important idea also with these sequences of word classes, is that there is structure there. It just happens to be that it’s not a tree like structure. And so, if we really want to understand how we use language, how language is represented, we need to start with how we actually sort of come into language, which is through conversing with others. It’s also presumably how language evolved by some ancestors of ours trying to communicate in some way face to face, using their hands, using their whatever sounds they could produce and so on. So, I think this is some of the big questions, and this is why I’m interested in this. And this is where this wonderful work with Yngwie is adding some really crucial pieces to this puzzle.
HEDVIG: I think that perspective, and this happens on our podcast quite often, we quite often talk to like usage-based linguists, and I think I’m on the same page as you are. I think the counterargument to that would be like: Oh, no, language evolved because we needed a way of structure our thought and our mental processes internally to ourselves. And that communication is not like its only, or maybe not even its primary function, in which case how it works in conversation is sort of like semi-uninteresting.
I’ve never fully understood that picture, and I believe I may be drawing a strawman here, maybe too negatively, but some people are interested in that perspective, and it generates very different kinds of outlooks. And this is why linguistics is so heterogeneous, I think, theoretically, because there’s so many ways of approaching even the basics of the field like what is language for?
MORTEN: I don’t think we want to exclude that language can be useful for structuring your thoughts. I think it can be. And certainly, for example, when I’m writing, I find that it actually helps discover things as I’m writing. So clearly, the use of language can be helpful in structuring your thoughts and so on, but that doesn’t mean that it originated for that. For me, at least, it’s like, yes, once we have language, yes, you can actually use it in that way as well.
I don’t think it’s a strawman. I mean, it’s sort of odd that there was a paper that came out just a couple of years ago in Nature that was a kind of stating this would by Ev Fedorenko and Ted Gibson and a few others [actually just Steven Piantadosi -D], if I remember correctly. And they’re sort of… Essentially, they’re stating that language is for communication. You would think that’s odd. I think for most ordinary people, why would it not be for communication? The fact that you have to have a paper in Nature stating that suggests that it’s not a strawman, right?
But I think for many of us, especially those of us who are interested in language evolution, that is sort of one of the key aspects of why we got language in first place. Now, again, obviously we could be wrong, but I think most of the evidence falls in favour of that, I think, currently, at least.
DANIEL: Can we go back for a second to the glue? I am a big fan of constructions as well. And I understand some constructions. Like, “peanut butter sandwich,” that’s a construction. There you go. You’ve got it. “I want,” I guess that’s kind of a construction as well. But I never really understood how once we had constructions, how do we link them together? At which point it just looks like trees all over again. But the sense that I’m getting from your discussion of glue is could it be that some constructions have overlapping elements that allow us to hook them together? Like, I guess we’ve got “to dance,” that’s a construction. Oh, “wannu,” that’s a construction. Ooh, they both have TO in them, I can drop those next to each other and overlap those. Now, I’m building something. Am I getting the right sort of picture?
YNGWIE: Yeah, I think you are. And perhaps I can start us off and then you can take over, Morten. Part of what’s happening here is, I think… So, the constructions you mentioned here are sort of intuitively, we can understand them as meaningful wholes in a sense. And I think that’s something that actually treelets does very well. It captures what we intuitively think of as meaningful wholes. On a side note, that might be why they are so prevalent. It’s because it’s understandable and easy to work with these meaningful wholes. And that might also be why it’s very hard to think of or to discover sort of any other kind of structure. [CHUCKLES]
But yeah, with your peanut butter sandwich example, there are some meaningful wholes to work with the construction, such as “peanut butter sandwich” or “I want,” or maybe even “I want a” if we are going back to the WANNA case, but then what? So, you must put them together in some sense. And I think that’s been a little challenging to specify in the constructionist literature, certainly because there are several ways of doing it. There’s the overlap that you talked about, but that’s not always the case. Sometimes, you can just put things next to each other. But then again, what is it that you’ve known, what is it that you’ve learned that sort of underlies this putting together, and I think this is where Morten’s glue is interesting.
MORTEN: Yeah. So, these non-constituent sequences, they allow you to combine constructions together, so you kind of need both. So, you have these constructions which are more traditional units that have been studied, but then we need ways of combining them together on the fly when we are coming up with what we want to say and that’s where these non-constituent sequences can come in, providing sort of a glue between the constructions.
Again, it’s a hypothesis. These are early days yet, but I think it’s an exciting hypothesis. At least, I think so. And it fits with this notion of trying to understand how we can use language as well as we do. I mean, it’s really, truly amazing that we can ever really use language given the time pressures and so on. So, I think that is really exciting and I think it’s often underappreciated just how amazing our language abilities are.
And I teach a class on the psychology of language and I try to get people to understand just how much their brains are doing, how clever they are when it comes to using language even though what they might be saying might not be so clever always. [HEDVIG CHUCKLES] But nonetheless, it’s an amazing ability. And figuring that out, I think we still have quite some way to go, which is good for Yngwie. So, he has lots of stuff to do. [LAUGHTER] But I think it’s exciting and it’s an exciting time to do this work, and I’m very proud of this work that Yngwie and I have done adding, if nothing else, sort of get people to discuss these issues, sort of not taking trees for granted.
HEDVIG: I agree. I also think it’s very exciting. And some of our conversations reminds me of a conversation with Dan Parker. But also, we talked to Steve Levinson again earlier this year about, like, how that sort of fascination and excitement about like how fantastic it is. I make little sounds with my vocal folds, and you get ideas in your head. That’s kind of bonkers. That’s kind of wild.
That said, and sorry to be the party pooper, but we don’t say very new things very often. I don’t know how to say this in a nice way, and I know we’ve talked about it before on the show, but people aren’t as innovative… if you gave a computer the grammar of English and the vocabulary of English and said, “Make up all the sentences,” and then you looked at all the sentences people use, I don’t think it’s all of them. People don’t say, “The transparent capybara stole my husband” very often.
MORTEN: Probably doesn’t happen very often either, but… [LAUGHTER]
DANIEL: Do we need that sentence? Really?
MORTEN: So, why is it that being a party pooper?
HEDVIG: Oh, because it sort of makes it a little bit less amazing, the fact that we can… We say similar things often, which means we can infer things and we can just use the same sentence we used before to say a new thing.
DANIEL: Well, it sounds like there are some new sentences flying around here in the last hour! [LAUGHTER] I’ve never said most of those things.
MORTEN: But I think actually, Hedvig, for me, that’s a positive thing because that’s part of what actually makes it’s possible for us to do what we are doing. So, if I continuously came up with new sentences that you never heard before, new ways of putting words together, that would just make it really, really hard on you. So, I think the fact that we are doing what you just mentioned, the fact that we don’t come up with new ways of saying things all the time… Now, we do come up with sort of… We might take existing chunks and put them together in slightly new ways, but we recognise most of those chunks, but that is actually what I think allows us to be both productive and talk about new things that we haven’t talked about, like some of the things we talked about today. But also, it’s actually what makes language possible in the first place.
And I think that’s sort of crucial. And you’re right that in everyday conversations, we use a fairly limited number of different words and a limited number of different constructions and so on, but that’s actually what makes it sort of possible for us to do it as quickly as we can. And I think we’ve been perhaps even sort of waylaid by this notion that there are all these new sentences that we’re saying all the time, and that’s what people have been trying to understand. Whereas really, what I’m interested in how it is that we can actually talk about the same things over and over again in real time, although it’s not exactly the same, but we still have to try to understand what it is. But I think that’s actually part of what makes it makes it possible for us to understand each other, so I think it’s a good thing. I’m the optimist sometimes. [LAUGHTER]
DANIEL: So, just to finish up, what do you both think needs to change about the way that we study language? What should we be doing differently?
MORTEN: Not look at trees? [LAUGHTER] Well, not take trees for granted.
YNGWIE: I like this “for granted” phrasing. I think it might be limiting our own scientific creativity a little bit. By taking it for granted, by assuming, “Well, this is what it is,” and then asking questions around that, how do I learn the trees? How do I use the trees? Not necessarily with trees. And I think there’s some interesting opportunities if you flip it around, if you say instead, “All right, well, I am learning. I am using. Okay, well, based on what I’m exposed to and my experiences, what might I learn? Might I learn the trees? Sure. But might I learn all these other alternative options?”
So, I like this way of sort of flipping it around a little bit instead of asking, “Well, how might we, from what we know of learning and so on, get to the trees?”, say, “Well, what might we instead arrive at if we say, ‘Well, here’s what we know about learning.’?”
MORTEN: I think that Yngwie just right on with what he’s saying. But I think also looking at memory, how is it that we can… What kind of memory processes… I mean, chunking is a memory process, what kind of other aspects of memory might be important for using language in real time like we’re doing now? But I think also more generally, I think one of the nice things about studying language over the last several decades is it’s become much more interdisciplinary. And so, we are beginning to realise that, for example, multimodality is an important aspect of using language, where that’s kind of being ignored before to a large degree.
But also, thinking about ecological validity, how we are using language in the real world rather than just always sort of focusing exclusively on the lab. Now, labs give us control and as a psychologist I am very much into control. But on the other hand, as a cognitive scientist, I like using computers, I like doing things in other ways and so on. And I think the field is definitely moving in that direction and I think that is a good thing. So, I think we’re on a good track, I think. But we have far to go. But that’s good because then we have something to do and it’s a lot of fun things to do.
HEDVIG: I think that’s a really nice note to leave it on because sometimes as a uni researcher you can feel that there’s some territoriality or that all good ideas have already been had and why should I even… This is like a depressive hole you can get into. I’m not saying that everyone has this experience. But what you’re saying here is like, “We made this neat study on like this kind of hypothesis on these kinds of informants.” I’m assuming you would like to see this replicated or reproduced on other datasets, on other conditions, other kinds of…
DANIEL: Languages.
HEDVIG: …non-constituent terms, other languages, other kinds of people, so there’s much more out there to find out that we’re only like at the tip of the iceberg, which is a nice… as a uni researcher, that’s nice to hear that there’s things left to do that Morten and every other senior didn’t just eat all the cake and then run away into retirement or whatever.
MORTEN: No, no, no. There’s much more to be done. So, you’ll have plenty of stuff to do. Don’t worry about that. But I think also in the same vein that too often, we forget about the history of studying language. So, I think we started today talking about some of the historical precedents and I think it’s important to keep that in mind because we… Not that they were necessarily right, but we can learn from what was done earlier. And I think too often, people don’t set the new findings that they have in a sort of a broader historical context because people have been sort of thinking about language for hundreds if not thousands of years. And yes, they didn’t know as much about the brain or how babies learn and so on back then, but they still had some interesting ideas.
And I oftentimes find it useful sort of looking back and finding something to build on and then sort of taking it forward and working with it with all my collaborators, which is sort of great. But I think, yes, there’s loads of stuff to do. So, don’t worry about this, Hedvig. You won’t run out of work to do.
DANIEL: And we won’t run out of show topics either. The paper is Evidence for the Evidence for the Representation of Non-hierarchical Structures in Language. It’s in Nature Human Behaviour. We’re going to slap a link up in the show notes for this episode. We’ve been talking to language researchers and linguistic legends, Yngwie Nielsen and Dr Morten Christiansen. Guys, thanks so much for coming on the show and talking us through your work.
MORTEN: Thanks for having us.
[BECAUSE LANGUAGE THEME]
DANIEL: We’ll do a proper outro on our next episode, but for now, big thanks to all of our patrons. Thanks for listening. We’ll catch you next time. Because Language.
IN UNISON: Pew, pew, pew.
[Transcript provided by SpeechDocs Podcast Transcription]