Caroline Rowland
@carorowland.bsky.social
Researcher in child language development @MPI for Psycholinguistics & Radboud Uni. She/her. "All models are wrong but some are useful" (George Box 1978). https://www.mpi.nl/people/rowland-caroline
End of a 3 year era. Kick off of our MEDAL final conference in beautiful Tartu... medal.ut.ee/event/final-...
October 9, 2025 at 6:44 AM
End of a 3 year era. Kick off of our MEDAL final conference in beautiful Tartu... medal.ut.ee/event/final-...
Unexpected benefit of working in a babylab
September 16, 2025 at 10:33 AM
Unexpected benefit of working in a babylab
Multilingual acquisition is one of the topics we mention in the Outstanding Questions section as something we need much more work on!
June 27, 2025 at 8:55 AM
Multilingual acquisition is one of the topics we mention in the Outstanding Questions section as something we need much more work on!
We also outline, in box 3, some ideas about why humans are so much better at learning language than large language models (teaser - unsurprisingly, its because LLMs don't implement all four components!) 18/
June 27, 2025 at 6:18 AM
We also outline, in box 3, some ideas about why humans are so much better at learning language than large language models (teaser - unsurprisingly, its because LLMs don't implement all four components!) 18/
But you haven't mentioned innateness at all, I hear you cry. Given that the question of what is innate has dominated our field for decades, why do you not mention this at all? Well we do - see box 2. But the jury is still out on that. We have some ideas - but those are for another paper :) 17/
![Text from the box:
Box 2. Types of constraint on learning
All theories need to specify what constraints the learner brings to the task. Contemporary theories acknowledge that there is no such thing as an unconstrained learner, but that is the limit of the consensus. Broadly, following Elman et al. [4], we describe constraints in terms of a 2D structure: domain-general versus domain-specific constraints on one axis, and structural (architectural and chronotopic) versus representational (knowledge based) constraints on the other (Figure I).
Figure I shows dimensions of constraints: horizontal and vertical two headed arrows; the horizontal arrow leads to the text representational on the left and structural on the right. The vertical arrow leads to the text domain general at the top and domain specific at the bottom.](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:iqz3cexpleulzmhlmugrputh/bafkreig4734mq4s7gwndo6h3ksf2r2ktxis64cdjueceglpt2gmhnjjzhm@jpeg)
June 27, 2025 at 6:13 AM
But you haven't mentioned innateness at all, I hear you cry. Given that the question of what is innate has dominated our field for decades, why do you not mention this at all? Well we do - see box 2. But the jury is still out on that. We have some ideas - but those are for another paper :) 17/
So that's it - four components that we think should form the skeleton of any future theory of development. Our claim is that "by embedding these four components in their own theories, researchers from every tradition are more likely to generate successful explanations of language acquisition'' 16/
June 27, 2025 at 6:02 AM
So that's it - four components that we think should form the skeleton of any future theory of development. Our claim is that "by embedding these four components in their own theories, researchers from every tradition are more likely to generate successful explanations of language acquisition'' 16/
4. Dyanmic developmental change. Development isn't the same as learning. Human development involves massive structural and representational change. And learning a language requires infants to build some of the most complex and abstract representations created by the human mind. 11/
June 27, 2025 at 5:51 AM
4. Dyanmic developmental change. Development isn't the same as learning. Human development involves massive structural and representational change. And learning a language requires infants to build some of the most complex and abstract representations created by the human mind. 11/
3. Active adaptive learning. This concept is fundamental to human learning. Unlike the way nearly all computer models learn, children adapt their attention and actions to maximize the affordances for learning in their communicative environment 7/
June 27, 2025 at 5:43 AM
3. Active adaptive learning. This concept is fundamental to human learning. Unlike the way nearly all computer models learn, children adapt their attention and actions to maximize the affordances for learning in their communicative environment 7/
2. Integration over multimodal input streams. Contra Chomsky the input is not impoverished but contains i) information on multiple levels within the speech/sign stream (e.g., frequency, prosody, phonology, meaning) and ii) multimodal information from visual, tactile, auditory & olfactory percepts 5/
June 27, 2025 at 5:37 AM
2. Integration over multimodal input streams. Contra Chomsky the input is not impoverished but contains i) information on multiple levels within the speech/sign stream (e.g., frequency, prosody, phonology, meaning) and ii) multimodal information from visual, tactile, auditory & olfactory percepts 5/
Now most modern models have structure-building mechanisms at their core. So though the idea of children *building* representational systems like language via interaction with environments was originally constructivist (thanks, Piaget) everyone now follows suit. But its worth saying it explicitly! 4/
June 27, 2025 at 5:33 AM
Now most modern models have structure-building mechanisms at their core. So though the idea of children *building* representational systems like language via interaction with environments was originally constructivist (thanks, Piaget) everyone now follows suit. But its worth saying it explicitly! 4/
So what are these 4 components?
1. Structure building learning mechanisms that create new linguistic representations by generalizing across individual percepts on the basis of common perceptual and functional features, and then storing the new representations in memory 3/
1. Structure building learning mechanisms that create new linguistic representations by generalizing across individual percepts on the basis of common perceptual and functional features, and then storing the new representations in memory 3/
June 27, 2025 at 5:30 AM
So what are these 4 components?
1. Structure building learning mechanisms that create new linguistic representations by generalizing across individual percepts on the basis of common perceptual and functional features, and then storing the new representations in memory 3/
1. Structure building learning mechanisms that create new linguistic representations by generalizing across individual percepts on the basis of common perceptual and functional features, and then storing the new representations in memory 3/
All four components derive from classic constructivist theorising - going right back to Piaget's original work 100 years ago. Which is why we call this framework for theory-building constructivist! 2/
June 27, 2025 at 5:21 AM
All four components derive from classic constructivist theorising - going right back to Piaget's original work 100 years ago. Which is why we call this framework for theory-building constructivist! 2/
One of the advantages of being an expat is appreciating the quirks of the language. This box is for foreign and old coins but the word for foreign is also the word for strange, so I always read it as "weird (and wonderful) old coins".
May 16, 2025 at 10:28 AM
One of the advantages of being an expat is appreciating the quirks of the language. This box is for foreign and old coins but the word for foreign is also the word for strange, so I always read it as "weird (and wonderful) old coins".
One of the best predatory journal emails yet. Not only wrong field, but also wrong name! Not that it's bad to be mixed up with my wonderful colleague Thea Cameron-Faulkner. But its quite a feat to mix us up - we even live in different countries.
May 13, 2025 at 9:44 AM
One of the best predatory journal emails yet. Not only wrong field, but also wrong name! Not that it's bad to be mixed up with my wonderful colleague Thea Cameron-Faulkner. But its quite a feat to mix us up - we even live in different countries.
This is very different to learning in LLMs like GPT 4, which passively absorb textual material and crunch distributional patterns. And partly explains why infants are so "cheap to run"; requiring far fewer resources. Infants are, put simply, the best, most efficient learners on the planet. 14/
April 25, 2025 at 4:43 AM
This is very different to learning in LLMs like GPT 4, which passively absorb textual material and crunch distributional patterns. And partly explains why infants are so "cheap to run"; requiring far fewer resources. Infants are, put simply, the best, most efficient learners on the planet. 14/
Instead, Jenni Sander has shown that children & adults use various non-verbal behaviours (touch,waving) to establish joint attention. She analysed head-mounted eye tracker recordings of interactions between 24 Dutch caregivers and 1-5 year old kids using spoken Dutch & Dutch sign language (NGT) 10/
April 25, 2025 at 4:34 AM
Instead, Jenni Sander has shown that children & adults use various non-verbal behaviours (touch,waving) to establish joint attention. She analysed head-mounted eye tracker recordings of interactions between 24 Dutch caregivers and 1-5 year old kids using spoken Dutch & Dutch sign language (NGT) 10/
In sign language interactions, visual attention to the object and visual language input (sign) compete for a child’s visual attention. So parallel input of object & language is harder to achieve. 9/
April 25, 2025 at 4:32 AM
In sign language interactions, visual attention to the object and visual language input (sign) compete for a child’s visual attention. So parallel input of object & language is harder to achieve. 9/
To better understand how caregivers and children jointly create interactional routines, Yayun Zhang set up a study using head-mounted eye-trackers with 16 parents & 18-24M children in USA, to identify where parents and children were looking and pointing, moment by moment during naming events 7/
April 25, 2025 at 4:22 AM
To better understand how caregivers and children jointly create interactional routines, Yayun Zhang set up a study using head-mounted eye-trackers with 16 parents & 18-24M children in USA, to identify where parents and children were looking and pointing, moment by moment during naming events 7/
In this way, they create ideal learning moments in interaction. E.g., shared book reading is a common daily activity for young children in Western societies. Shared reading integrates speech, gestures, visual attention & environmental context, making it an inherently rich & multimodal experience 6/
April 25, 2025 at 4:19 AM
In this way, they create ideal learning moments in interaction. E.g., shared book reading is a common daily activity for young children in Western societies. Shared reading integrates speech, gestures, visual attention & environmental context, making it an inherently rich & multimodal experience 6/
In our commentary, we completely agree, but add information from a different perspective. We contend that, to understand multimodal language acquisition, you have to acknowledge that it takes place in *social interaction*, in which every single step is multimodal, in every language modality. 3/
April 25, 2025 at 4:14 AM
In our commentary, we completely agree, but add information from a different perspective. We contend that, to understand multimodal language acquisition, you have to acknowledge that it takes place in *social interaction*, in which every single step is multimodal, in every language modality. 3/
Well, we don’t think it is because the gesture scale is measuring something different to the vocabulary scales. In fact this is almost certainly not the case – there is a lot of evidence for a close language/gesture link, due to the common symbolic nature of both language and gesture. 11/
April 8, 2025 at 1:49 PM
Well, we don’t think it is because the gesture scale is measuring something different to the vocabulary scales. In fact this is almost certainly not the case – there is a lot of evidence for a close language/gesture link, due to the common symbolic nature of both language and gesture. 11/
So what is going on here? Why does gesture seem to be better associated with language-related risk scores than the measures of language themselves? 10/
April 8, 2025 at 1:46 PM
So what is going on here? Why does gesture seem to be better associated with language-related risk scores than the measures of language themselves? 10/
A loooooonnng time ago, Katie Alcock, Kerstin Meints and I created and normed the UK-CDI Words & Gestures. CDIs are parent report inventories of children’s language, in which parents tick off the words and gestures their children comprehend and/or use. 2/.
April 8, 2025 at 1:36 PM
A loooooonnng time ago, Katie Alcock, Kerstin Meints and I created and normed the UK-CDI Words & Gestures. CDIs are parent report inventories of children’s language, in which parents tick off the words and gestures their children comprehend and/or use. 2/.
1st results from the 1st ever Sign Language of the Netherlands CDI. Results for children's comprehension and production, reliably correlate with age. All the work of the extraordinary Jenni Sander and Dilys Eikelboom @jennisander.bsky.social
April 5, 2025 at 9:08 AM
1st results from the 1st ever Sign Language of the Netherlands CDI. Results for children's comprehension and production, reliably correlate with age. All the work of the extraordinary Jenni Sander and Dilys Eikelboom @jennisander.bsky.social