 Okay, welcome everyone. My name is Joey Lovestrand. I'm a British Academy postdoctoral fellow at SOAS University of London hosting this webinar on behalf of the linguistics department. Our speaker today is Dr. Limor Raviv, who is a postdoctoral researcher at the Artificial Intelligence Lab at Free University of Brussels. Limor also soon start a Minerva research group called Language Evolution and Adaptation in Diverse Situations within the Max Planck groups. And we'll be speaking to us today about the links between language evolution, language acquisition, and language diversity. So I'm going to hand it over to Limor to give us a presentation, which will last 30 or 40 minutes, then we'll use the rest of our hour together for questions and responses and so at that time you can feel free to put your questions in the chat or raise your hand or otherwise signal and ask a question directly. So thank you to everyone for joining us. Thank you, Limor, for preparing this presentation and look forward to hearing what you have to share. Thank you. Okay, I'm going to go ahead and share my screen. And hopefully, yeah, can you see my screen? Great. Yes. Okay. So today in this talk, I'm going to give you a taste of my research in the past six years, which attempted to link these three topics, language change or evolution, language acquisition and language diversity. And it's a bit of an overview. There's a lot of projects I'm not going to mention, but maybe you can ask about them later if you want to or if you already have any clarification questions in the chat and you think it's really important for me to answer them now before I move on. You can also unmute yourself and maybe just draw my attention because I can't see the chat while I'm giving presentation. So I always start my talks with the same slide. I've been using this slide for the past few years. And it's this one. It's because to me, one of the most interesting questions in the field of linguistics are why are there so many different languages in the world. And of course, in this slide, you only see kind of like the major big known languages in the world. But actually there's about 7000 different languages. Some of them are very small. Some of them are in the risk of extinction. And I wonder what are the possible sources for this astonishing linguistic diversity. And I'm also curious in how different languages evolve. So what are the cognitive and communicative pressures that shape the evolution of language in our species and also in others. And finally, why do languages differ so much from each other and how do they differ from each other and are some languages easier to learn than others, what kind of mechanisms enable language learning across childhood and across modalities and so on. These are a lot of big questions. And in my research so far I've been trying to shed light on these questions using kind of small pieces to the puzzle and using experimental methods and more recently computational methods. And I'm going to show some of this work to me. Okay, so I think the first thing to consider when talking about languages is that languages are constantly changing. And certainly if there is one single universal pattern that we know to be 100% true for all languages in the world and really there's only one of these universal patterns is that if people use these language to communicate with each other, these languages will change over time. And we can think of language evolution and language changes the process in which linguistic innovations or or variants emerge and then spread. In the case of the cultural evolution of language the story is quite simple. It's this. Well, people they they innovate all the time. They randomly or intentionally produce new sound variants they create new lexical words we use new grammatical But only some of these variants are then adopted by others and spread to the entire community. So this process is generally termed cultural evolution. It's somewhat analogous to biological evolution, in the sense that you have mutations and they spread as a factor of natural selection so some variants might be better than others but also a lot of random changes and random. But I'm interested in how this process actually occurs in languages. So how do language innovations emerge and spread and how fast can they become the new linguistic one. So to do this to answer this question I'm currently using agent based computational modeling. In these models you take different populations of simulated computer agents, which have a carefully controlled capacities, and each of them use different variants in this illustration here, every circle is a person, their color is what variant they use and you can see the connections indicating who speaks to him. And then you can examine how different variants tend to spread in a population over time, until most agents are using the same model as you can say that a norm has been established. But more importantly, we can use these models to examine how this process is affected by the cognitive capacities and the communication patterns of these agents. And this is done by manipulating things like how flexible or stubborn these agents are. So how likely are agents to adopt to other people's variants or how likely they are to stick to their own unique dialects and forms, or do agents have a preference for some variants over others. How many variants can they store in their memory. And who do they tend to copy from so do they tend to copy from the person that is most prestigious or most connected in their community or rather their most closer kids or friends. These types of computer models can provide really valuable insights because they break down a very complex process of language change into hopefully small, transparent steps. So just to give you a little taste of that. And some simulations that I've done, we looked at agents personalities. So what I said before how likely they are to adapt or to accommodate to others. And we checked whether the proportion of flexible and stubborn agents affect convergence. And the results of this different simulations might not surprise you they might be a bit trivial, but it shows that it doesn't really have a big impact on language change. So even if there is a lot of stubborn agents and people are very conformist and they don't really like to change. This only slows down the spread of variants the entire community so it slows down language change but it doesn't stop it. So as long as agents are not 100% stubborn, and they do accommodate a little bit. Sometimes, the network will eventually converge on your variants and this is important because I think it shows or why indeed it is the case that all languages change all the time even if we don't want them to they just do. And I'm going to get back to some more computationally computational modeling work later on. But now I'd like to shift the focus to some more experimental work I've done, because I think that we need to take a step back before we talk about this kind of language change. So indeed the first important question to ask is how do people even learn the variants of their language and looking at language acquisition one of the biggest questions in the study of human language is how do children really discover the structure of their language, and more generally how do they learn repeating patterns and regularities in any form in their environment. One possible answer is that children have what is called statistical learning abilities. So this is the ability to implicitly detect patterns and sensory input, based on distributional properties statistical properties, things like co occurrences and transitional probabilities. And in the first behavioral study I ever conducted I did this at the very beginning of my masters. I looked at this crucial mechanism and I asked two questions. The first questions was the first question was whether statistical learning abilities are stable across development, or do they change with age. So is it the case that you learn better or worse as you grow up. The second question was whether these learning abilities are stable across sensory modalities so our learning outcomes the same, or it doesn't develop differently individual domain or the auditory. To address these questions I conducted this large scale study of statistical learning across modalities and across development. I just 230 children between the ages of five to 12 on matching visual and auditory tasks and we looked at how well they learned the patterns in their input. So in this plot that you see here you have on the x axis the age of the child, and on the y axis how well they learned and the color coding is for whether it was visual or an auditory task. So if we compare the children's accuracy on both tasks we found a difference in the developmental trajectory of statistical learning across modalities. So although learning improved with age individual modality. It was age invariant in the auditory modality. This means that auditory learning just did not change much over childhood, while people got better at learning children got better at learning visual patterns as they grew old. So the statistical learning develops differently across domains and suggest that this is not a unitary and stable capacity but rather modality sense. Okay. So next I want to study language evolution in the lab. And specifically I wanted to look at the emergence of linguistic structure in the process of cultural transmission in a lab setting. This is using iterated learning experiments where participants learn a target behavior, like an artificial language. And then their learning output is given as input to the next person. So what this means is that you come to the lab, you learn an artificial language or made up mini language for something, you get tested on this language. And then whatever responses you put in the text and in the test box are then given as the input language to the next person. So this is like a broken telephone game where you see how the changes occur over time over a generation or chains of participants and previous work by Simon Kirby and other colleagues at Edinburgh showed that this process can really lead to the creation of grammar, basically systematic compositional linguistic structure over time. So this effectively shifts the explanatory burden for the origin of grammar from having some biologically evolved capacity to just having a cultural process, the shapes linguistic patterns according to our cognitive needs and biases. But when I looked at these studies I was a bit concerned because I kept thinking, hey, what about children, right because one important criticism of these types of studies is that everybody who do these studies all the participants, and these learning tasks are adults who already speak a language like us. So the problem with this is that older participants may benefit from better working memory more sophisticated problem solving strategies. We might rely on our previous very extensive and explicit linguistic experience. And if this is true then maybe the findings from previous iterated learning studies are not necessarily evidence from some underlying process that's responsible for the evolution of language but rather just the result of already knowing a language than bringing all your biases to these tasks. So I thought it is really crucial to test this paradigm with children as well, because children are indeed the main language learners in the real world. And they also have much less explicit knowledge of the grammar of their language so they're much less likely to do this kind of grammar introducing on purpose. So what I did is I created this kind of child friendly iterated learning experiment and I compared the performance of 140 children to that of 100 adults in two different conditions, like in the original study, we basically tried to replicate this but with children. And here what I'm uploading on the X axis is time so basically generation number how many people have passed through this chain, and on the Y axis it's the structure score so how systematic how grammatical the languages ended up becoming. And when we compare the performance of children and adults in terms of their transmission error and their creation of structure, what I found is that adults consistently create compositional languages they significantly outperform children they're just much better at these traditional language learning tasks, and children's languages didn't show a significant increase in compositionality. You can see this because there's a black line, and that black line is chance so basically everything below this black line didn't reach a threshold of what we would call a grammar, and everything above it would, you can see that the adults go over it, but the children don't. However, I want to also draw your attention to the fact that in children's languages did become easy to learn in the same way as adults. And some children did develop languages with significant systematic structure, they did this on multiple occasions you can see some blue lines kind of popping up, but it didn't survive the transmission process this means that the next kid that didn't pick it up. And I think even more importantly, is that we, we use additional analysis to show that the difference between children and adults in this paradigm was related to how well they learned their input language. We're linking these findings back to language acquisition. What we found is that participants didn't matter if they were children or adults who learned the input language is better, we're also more likely to introduce more structure to their output. And this kind of strengthens the idea that you have to first learn your language and have some biases of your language in order to introduce or regularize any structure. Okay, but next I wanted to study this process of language evolution in a more realistic setting, because in the real world, clearly people don't just learn and transmit languages from generation to generations and isolation. Right, we actually interact with other people. We are subjective to communicative pressure when interacting. So I wanted to test how languages can evolve in a community from scratch via communication. And all of this grammatical structure emerged because of interaction. So to do this, I developed this group communication paradigm for testing the emergence of structure and spread in a miniature community. So what I did was I brought participants to the lab and I asked them over the course of several hours to create a new artificial language to communicate with each other. Participants in the same group are paired with a different person every time they play a bit and then they move and talk to another person in their group. And they need to interact about different types of novel scenes. I'll show you how these look in a second basically they're just different shapes that are moving on the screen in different directions. And they earn points when they successfully understand each other. Most importantly, these participants are not allowed to use Dutch, which is their native language or English or any other language they know. They can't point, they can't just do nothing. They need to come up with kind of nonsense gibberish words in order to describe these scenes. And I want to show you how a typical interaction looks like just so you kind of get an idea of how the game is working. So in a single interaction, one participant who is the producer sees a shape moving on the screen in a given direction, and they need to provide a label for it. So let's say if the participant decided to call this item WAPE, if this word is legal, that is it's not in Dutch or English, we pass it to their partner, the guesser. So the guesser sees this word WAPE together with a grid of eight different items and they need to select what they think their partner meant. And I just want to reassure you that although this looks very epileptic and pretty old like a lot going on, our participants get used to this very quickly. So over a few minutes already, they can tell apart the shapes in there, and they just do this again and again that it becomes very normal. Okay, and once the participant clicks, which shape they think WAPE means, both of the participants get feedback. So they can see which was the right item and what's their partner selected and they can learn from it for the next interaction. And then these two participants, they switch roles. So they alternate between guessing and producing, guessing and producing, guessing and producing multiple times. And after they do this for about 20 minutes, they leave and they go interact with somebody else in there. Okay. And importantly, at the beginning of this experiment, the participants are really, really guessing what they have no idea what's going on. They're making up random names. They're just clicking on things. But over the course of several hours, they slowly start developing linguistic structure and regularities, which are anchored in some successful interactions. So essentially, they start creating a grammar. And in this grammar, there is a different part word for describing the shapes and different part words for describing their motion. And we see this again and again and again in this group communication paradigm. Different groups create different grammars, but essentially they all kind of converge on this idea of having some systematicity in their lexicon. And through this experiment, we can examine these emerging languages on several measures. We can look at whether participants successfully understand each other. Whether the group ends up converging on a shared language or is it just a case that everybody just stick to their own language. You can look at how much do these languages change over time and how systematic and compositional these grammars are. And this paradigm really allows us to look at how languages evolve in a community in real time. And to see that like very these kind of basic simple communication pressures can really between people can really lead to the creation of systematic grammars over time. Which again, I think is really cool. Also, then they spread to the entire community. Okay. So this work on language evolution led me to ask questions about language diversity. So why do we have so many different languages in the world with so many different levels of grammatical complexity. So one thing that may cause languages to be different is the fact that they're spoken in different social environments. So languages evolve in different communities with different population sizes, different social structure and different social needs. And the idea is that languages may adapt to fit these different social needs. So using the group communication paradigm and just showed you, I was able to experimentally tease apart different social features that may affect language diversity, namely community size and network structure. And to test how they shape the evolution of languages in the lab. So focusing now on the effect of community size. Here I asked how does changing the size of the group the size of the community affect the languages that would evolve in it. And how does interacting with more people lead to a difference in the languages structure. So to this end I compared languages that were formed in small groups of four people to those formed in larger groups of eight people. And I want to reassure you that even though it's clear that both of these groups sizes are really small compared to the real world and the real world even a group of eight people is very, very, very, very, very, very small for our language community. But the point is that in this miniature setting in an artificial language setting where you have only a few meanings to communicate with and only a few hours to do so, just doubling the number of people in the group already makes a significant difference to the languages. So it's a miniature setup and also a miniature manipulation, but doubling the number of people has an impact on these languages. So let me show you what we found. Again on the x-axis and plotting time. So this is the time since the beginning of the experiment on the y-axis and plotting how systematic the language is. I'm not going to get into how we measure did this now because it's a bit complicated but if you're interested in knowing yeah how do you measure grammar in these languages then please ask me at the end of the talk I'm happy to explain. And what we found here is that languages of big groups were more systematic and the big groups created this more systematic grammar faster and more consistently. So I don't know if you can see the little lines behind the very thick line so every line represents a group. And what you can see is that the small groups the blue lines, some of them are really high on the structure but some of them are really low so they're kind of all over the place. But if you look at the big groups, the red lines, they're all kind of very systematically and consistently developing higher levels of structure together. And the explanation for this trend was that the need to develop systematic languages is much higher in a larger. So the idea is that people in larger communities have less shared history with each other, and they're exposed to more input variability just more variability in general. So maybe in a small group you could just remember each other's unique variations. I can remember that you say booba and you say kiki and that's fine and I can just remember it. But this is really hard to do in a big community. You have much more people to interact with. There's a lot of variability and therefore there is somehow a pressure to favor something that is simpler and more systematic more grammatical that can help you ease convergence. So although convergence is harder in a big group, just because it's harder to agree with eight people than with four, it's also much more need. Okay, and this means that as I said members of the big communities are seemingly under this pressure to create languages that are more generalizable somehow that can help them to communicate. And these experimental findings showed the community size can really affect patterns of language diversity, such that the social and communicative pressures that are associated with language use can shape them really the nature of languages created in different communities. Okay, but this also has conclusions, really important conclusions for language learning language acquisition, because think about it, if languages with more regular and compositional grammars are easier to learn, as is assumed by many, many, many people. And if such languages tend to develop more in big communities, and this suggests that some languages are acquired faster than others. And that this learning advantage can then be traced back to community size and to the degree of systematicity in the language. And I have to say this goes against a really widespread axiom in the field of linguistics at least back in the day. Let's say that all languages are equally hard to learn and take the same effort to acquire. There's a lot of kind of recent evidence, even from child language acquisition that shows that this is probably not the case. Some languages, for example, Danish seem to be much harder for children to learn that even when you compare children, for example, at the age of five, their linguistic knowledge and their grammatical knowledge is kind of what we would call delayed compared to other languages like English, which is considered really simple. And this is this is something to keep in mind and I have to say when we use terms like simple complex regular. I see here my co-author Cedric in the audience and I'm just going to mention this. There is no, there's nothing fundamentally better or worse about being simpler or being more complex. It's associated with different things. For example, a simple kind of clear transparent predictable language can be really, really good for learning really, really good with community for communication and a complex language. Some people might think, oh, it's more sophisticated it's it's more elaborate, but those are just judgments that we kind of put on these terms, they mean nothing, except for a description of how systematically and transparent these grammars are. I'm just writing a piece about this point right now. But again, so this is just about how easy it is to learn and to use. Okay. And so to test this idea that more structured languages are easier to learn. We conducted another study basically what I did is to take the languages the final languages that were created by the big and small groups in this previous experiment I just showed you these languages varied and how structured they were. And then I just taught them to 100 new people that were not a part of the original community. So these people were exposed to the language and they had all these trials and training blocks where they had to guess the words and produce the words and they were trained on these words for about an hour. And then we gave them a test to see how well they learned. And the results were very clear. They showed that languages with more systematic grammars again here it's loaded on the x axis how systematic the language is on the y axis how well people learned it and reproduced it. So languages with more systematic structure were learned better participants were more accurate at reproducing these languages they learned them faster and they learned them more consistently. Again, I just want to point out. If you can see, I don't know if you can see my cursor. I'm going to hope you can so on the very left edge of the of the plot are languages with very low structures languages that are not structured at all they basically don't have a grammar every word represents something else and you just have to memorize all of them. As you can see, some people were really good at learning that that wasn't a problem they were able to learn it, but some people were really bad. Okay, so the individual differences were really big for the small language for the unstructured languages. But when you look at the highly systematic languages, everybody is clustered at the top everybody can learn it everybody reaches ceiling. So it doesn't really matter if you have good memory if you're very motivated, you can learn it easily in one hour. So this is really important because it confirms this relationship between learnability and systematic and even more interestingly, I wanted to know how can people generalize the language they learned to new unfamiliar meanings that they've never seen before. So, can learners actually apply the languages they learned to new contexts right that's what makes our languages so productive and useful, and also will different people will different learners do this in the same way. So to this end, I gave participants another test at the end. They saw 12 scenes with new combinations of shape and direction that they haven't been trained on before. And they were asked to label them according to what they thought would be correct based on the language they learned. So participants in this part could do whatever they want. They can make up totally new words so they don't generalize at all. They could use homonyms. So replicates words and reuse words they've already learned. They could combine words they can use the rules of the language if there were any they really could do whatever they wanted. And more importantly, I looked at the generalizations of different participants that learned the same language and asked how similar they are. So do different people generalize new scenes in the same way and use the same words, despite never learning these words, never interacting with each other. So when we look at the similarities between these generalizations, again, very clear results. Participants who learned more structured languages were much more likely to produce the exact same words, the same generalizations. And this finding suggests that systematicity can allow strangers to converge effortlessly. So people who never interacted before could potentially communicate successfully about new things and immediately be understood. And if you remember, this is exactly the mechanism that we postulated for the group communication paradigm that I just showed you. So it's about having more structured languages to facilitate convergence in a big group with more people. And this finding here directly supports this idea. It shows that the benefits of linguistic structure or grammar go beyond learnability. It's not just that it's easier to learn. It's also that it's advantageous for communication between the individuals and it aids productivity and general languages. Okay, so the next thing I did was look at network structure. So does it have the same effect as community sites. And here I asked what happens when groups are of the same size, but they vary in how much and how community members, community members are connected to each other. So this is, you know, in the process of language evolution, can we also say that it's shaped by the density or the sparsity of the group. And in the previous experiment, all the groups were fully connected. So everybody in the group spoke to everybody else in the group. But in the real world, this is rarely the case, right. Larger communities are typically more sparse less connected. Many people never meet you on some people in the community you will never interact with. And also now everybody is equally connected. So some people interact with many other people, and other people are more isolated. So looking at the role of network structure and density independently from community sites. I went and compared three different types of networks three different types of groups. The first group was this fully connected group in red. This is exactly like before. So everybody talks to everyone. So although this resembles early human societies, it may be quite rare nowadays. So maybe we can see this in some hunter gather communities or some villages, but it's overall pretty are pretty rare network structure to have. The second network we looked at was a small world network. This is here in blue. These networks are much sparser. They have much less connections in fact, exactly half so we delete half the number of potential connections between people. They are more realistic in the sense that strangers here are indirectly connected. So, for example, participant G and H they never interact directly. You see that there's no line between them, but they are connected indirectly via participant F for example participant D. So innovations can still flow in this community. And the last network we looked at was a scale free network in green. So these networks have been argued to be the most representative of modern human societies. They're also sparser. They're exactly like the blue networks in the sense of having the same number of connections. But they follow what is called a power law, where few agents are highly connected. In this case, if you look at participant A, he is what we call the hub. He is connected to almost everybody else in the group. But most agents are less connected. For example, participant D or D, which are quite isolated. Okay. And here our prediction was that the sparser networks, the blue and the green, would create more systematic languages for the same reasons as before. Because sparser networks are typically more variable and people have less shared history and it's hard to remember the variants of all these people that you may not interact with directly. Convergence should be harder in these networks and therefore much more needed. So they might be under a stronger pressure to create systematic languages. But spoiler alert, this is not what we found. So when we looked at the results of this experiment, we found no significant differences between these three network conditions. So all the networks showed the same degree of high structure throughout the experiment and also showed similar behavioral patterns when looking at all the other measures I mentioned before like stability and accuracy and so on. So you can really see these lines are all kind of going together in the same way. And in retrospect, this result makes sense, because we found that there was no differences in the input variability across these conditions. So it was not the case that the sparser networks were more diversified. But I do want to urge you that although this doesn't seem to have an effect in this experiment. So when sizes kept constant and relatively small, it's important not to draw any strong conclusions from this null results. That's eventually what this is. There's many reasons why we can imagine that network structure didn't have an effect in this current design. Maybe our network didn't differ sufficiently from each other or maybe they were not big enough, maybe in a network of 20 people that would have worked. Or because of other mythological limitations and in the discussion of this paper here we actually list a few things that we really think could be improved in future research so we really need to take another step in order to understand what is the true role of network structure. So this is kind of until now was kind of an overview of what I've done so far. But I also want to give you a taste of glimpse into some of my future projects because soon as, as you heard from Joey I'm going to start my own research group at the Max Planck Institute, a group called leads in short, and I have three major projects and I want to share them with you. So the first project is about linking experimental results to data from computational modeling. So the idea is to take the pattern that were obtained from real groups of human participants during the communication, the group communication study I just showed you, and to compare them to patterns obtained from different sized populations of these people using simulated agents. But in this case the simulations, the simulated agents are not kind of too simple in order to really match human behavior. So we're going to use deep learning. And this is a collaboration with other brilliant scientists. The idea is we're in very, very early stages of this project. We're going to develop a very close model of participants behavior using groups of networks of interacting AI so neural networks that will interact with each other and do the same thing that our participants are doing. So this is the architecture of the network it doesn't really matter the idea is that it will have different levels. The first level will teach these networks how to see the world so like our participants they first need to learn how to differentiate between these different shapes moving in different directions and to identify. These are the relevant features for categorization. The next thing is to match these scenes just like the participants are doing. Then there will be, they will learn slowly how to speak and listen to how to associate a string of text to to these scenes that they've just learned how to see and categorize. Well, and I think that's that's really important to learn how to speak and listen. And this kind of brings us back to questions about language acquisition because how will these different neural networks learn to speak and listen. How fast can they pick up on the structure and regularities of their environment. This is very likely to be through the same processes that we've mentioned at the beginning of the talk like statistical learning or Bayesian inference which is what our computer models are doing. So tracking the core currencies and updating the possibility of a given mapping as a result of experience. And once we have that established we can return to examine the evolution of grammatical structure and to this and we plan to teach these networks how to successfully communicate with each other. So basically having different networks interact such that one network would produce a text for a target scene. The other network would take this text and need to match it to the right scene from a set of possible scenes. Really exactly the same type of interaction that we had in the experiment that just showed you with human participants. So the question is when these networks interact with each other over time. Will their representations and associated text also changes a function of interaction. And will we see the same patterns of emerging structure and emerging grammar in a population of AI. And of course we can also ask questions about language diversity so manipulating the size of this AI population and see if we can replicate the experimental findings that larger populations develop more structure languages. Okay, so this is kind of this is one project. The second project. The second thing we can do with kind of linking computational modeling to this group communication paradigm is to investigate new and exciting topics about community structure. And here in the second project we're going to look at age and gender diversity and how they shape the process of language formation and innovation. So one of the most intriguing findings in social linguistics is that the gender and age of speakers can really affect the process of language change. Specifically was suggested that women are actually the leaders of language change. There are some studies that show that women are often more linguistically innovative than men. They tend to adopt new variants more frequently. And similarly there are studies that show that teenagers around the age of 17 also tend to create and adopt new variants much more often than adults. And although I think we have all of an intuition as to teenagers changing languages and having their all kind of new slang, why women would be leaders of language changes a bit less known. And more importantly, all these claims are based on several case studies. So all we can do in these kind of documented language changes to see a language change happening in process, and then try to trace it back to where it started but once we are investigating it, it's already very much in the process of change if we've noticed it. So it was never tested experimentally or computationally from the very, very, very early stages of emergence in real time, so not retroactively. So we currently don't know how population differences in gender and age really influence cultural evolution, language innovation and language change. So the linguistic tendency is really going to be associated with male versus female speakers and adult versus teachers. So in this future project, we will couple the experimental methods and computational modeling to determine whether women and adolescents really lead the process of language change. Specifically, we will use this group communication paradigm and simulated populations to manipulate the composition of these many societies in terms of the gender and age of individuals. So for example, testing different groups of participants that are either homogeneously men here in blue, men biased, gender balance, so half-half, women biased or homogeneously women. So or in the same for age, so whether groups of all teenagers, all adults age balanced and so on. So real such changes in the identity of group members influence the creation and spread of linguistic norms over time. And is it really the case that women and teenagers lead this process and adopt and create novel words faster. Okay, and last but not least, and these are my few final slides. So the next project is about looking at language evolution and adaptation to a virtual reality environment. So here what we're going to do is virtual reality environment offers high ecological validity and a more naturalistic setup that allows you to have still high experimental control. So we already have the stimuli for this experiment. It's going to be different kind of funny novel creatures that are different their shape their size their movement types or whether they hop or bounce and their movement speed. So the goal of this adaptation is to test how modality differences affect the process of language evolution. So specifically, we're going to look at the process of language formation when using gestures. When using vocalizations or when using book and looking at how gestural and vocal languages are formed in the VR environment we can really see the multimodal origin of languages. So if people wonder how did the first languages look like where they're signed where they vocal. When did we shift a lot of people think that we started by gesturing and then shift to a vocal languages but why. So what are the comparative advantages and disadvantages of each of these modalities during the emergence of a novel communication system. So let me just predict for example that grammatical structure and iconicity will emerge in both modalities. You can imagine size the size of the creature being expressed iconically in the signed modality by using a bigger gesture, but it can also be expressed iconically in speech by using a high or low vowel for big and for small and big items. But it's very likely that there would be more systematicity or more iconicity sorry and the gestural modality because it's easier to have iconicity and gestured modality. And more importantly we can use this VR paradigm to look at questions related to language diversity and specifically asking if languages that also adapt to their physical environment. What happens if we add noise and these experiments would adaptations occur. In particular we plan to introduce wind to the VR environment by playing a sound on the speakers that masks the s sound. And our prediction is that in such a windy environment, the vocal languages would include less s sounds or even not at all, because these would be much harder to detect so really showing that the conditions of the environmental conditions can change the phonology or the basic building blocks of language. And then the gestural condition we will introduce poor lighting so basically dark in the room. And in this case we predict that the sign languages will also have to adapt by making the gestures larger and more salient, maybe accompanying them with vocalizations. And this can also influence the interplay between iconicity and systematicity because if larger gestures can no longer be used to indicate larger creatures, this might force participants to shift away from iconicity into more arbitrary gestures. Okay, so that's it for me I just want to conclude by saying that my research tried to tie these things using using different methods and I want to uncover the time course and social dynamics and cognitive mechanisms that characterize these processes. There's a lot of other projects that didn't talk about, for example, looking at complex cultures of elephants. They're all of children and language change. I have some swarm robotics models that we use a lot of stuff, but they all share the same goal and that is to promote this kind of multifaceted understanding of a very complex phenomenon. And if you have any questions I'm very very happy to take them now and I just want to thank you for your attention. Thank you so much. It's a really fascinating presentation switch my view here. So do feel free to put your questions in the chat or raise your hand if you like your question we did get one question already from Abu. So if you have any comments on bilingual multilingual people so I guess that's one more, you know, complex factor is what happens with language context. It's understandable why you started this and more of a monolingual context but what do you think what would you think in terms of how that plays into this whole story. Yeah, that's a great question. In a way, none of our participants are really monolingual in the very pure sense of the word they are at least. Some of my participants were recruited in Israel, like the first experiments I showed and then they speak Hebrew and often English. And in the recent work it's Dutch speakers who often also speak English so in a way everybody is a bilingual and everybody's a bilingual when they come to learn the artificial languages right they all already know the language. And now they're learning this in a new context. And this definitely might have an effect because I think prior knowledge is really big determined, really big factor in how well and how easy it is to learn a new language. So we find that you know it's easy on some levels it's easier to learn a language that's similar to our languages we already know in terms of structure and grammar. On other levels it's harder because it means that there's more competition, some words are very similar and only a tiny difference and maybe sometimes it's also easy to learn something that's remote. And there's a lot of research on what makes second language learning easier or harder and as we said similarity is a really big factor but also motivation. I think, for me, those questions are only interesting, interesting as far as they can tell us something about general language use so that the nitty gritty details between L one and L two and whether it's different because of this morpheme or that morpheme. That's very interesting for many people who work on language learning but for me, interested in kind of the big giant picture why do our species have language and how does it look like. It's not only matters because prior knowledge I think is what tells a part, not only bilinguals to monolinguals but also children and adults. Also, people with late exposure to language and early exposure to language and those are kind of the, that's the, that's where the, the gold consequences are it's, when does it, when does it break down learning, or when does it boost something for the reason I went to study children and the very first experiment with iterated learning is because I was afraid that exactly being a proficient bilingual or in this case a proficient language user might really change the patterns that we see in this experiment. And I was happy to find out that with children on some levels, even though they had less knowledge and they were generally pretty bad at this experiment. It seems like they're going through the same patterns and the same. Yeah, the same trajectories that you would expect from a, let's say a poor, discolored adult. And that's interesting because it means that these paradigms can tell us something valuable about language evolution and language change. And if children would not be able to do anything similar to the adults, I would be in serious doubt in taking any big conclusion from these iterated learning models because I'd say okay if they don't work with a non-proficient language learner, then they're not very interesting. But I hope that answers your question. Jonathan, you had a question? Yeah, I do. Yeah. So, Leymour, thanks very much for a really super interesting paper. You're asking some massive questions and, you know, trying to innovate some really interesting research designs to get at them. So, you know, I can only applaud that. I was kind of, I sat there reflecting on your social network experiment for a little while. And I was trying to work out, you know, what might have gone wrong and why you had to sort of non-significant results. And then you started talking about the social linguistic literature, right. Correct me at any stage here or feel free to shoot me down by all means I'm just sort of, you know, spitballing. But, you know, when you started talking about these sort of gender effects, for instance, you know, they're couched in a very narrowly specified theory that have not been tested, you know, certainly outside of big dominant Western languages, right. And there's plenty of data now to suggest that elsewhere you don't find these patterns. So I kind of wondered at that point, whether this is related to the social network experiment in that it's difficult to capture, right. Cross social cross cultural intricacies in this modeling and I just want I just, I guess I'm just offering some reflection and wondered what you have to say about that. It's right on the money because one of the reasons we're doing these experimental work and this computational models is to really tease apart and try to isolate this one specific thing that may cause languages to be different, which in the world in the real world, even when we find a pattern, we don't know what to attribute it to it's correlated with gender. Great. But that doesn't say anything about causality. And it doesn't say anything about why gender is driving this effect or what is it about gender. Maybe there's a mediating factor altogether. So indeed, for the experiments with the age and gender, I was just presenting them very briefly, but the goal is to, in a way, dig deeper. So if we do find that, for example, adolescents in this paradigm tend to innovate more, it will be very interesting to find is this something about their social role in society. So the fact that they're less pressured to conform, for example, or that there it can even be something about, you know, hormones or peer pressure peer to peer convergence. Those are and then you can try if you have a good theory about why that would be the case you can simulate it in these experiments with not adolescents. So take adults, put them in different social niches where one has to be conformist or prime them with flexibility patterns or or or innovative patterns, and then see whether you get them to behave like adolescents in this paradigm. So uncovering what is it about this age group that drives this effect in the same for women because I think the literature about women as leaders of language change was first of all it's rather some of it is new like you know work by by Eckhart showing you know constructions like like and be like they come from teenage girls. There's also some work in from the 70s about other more isolated languages. And I think here the question is what is it about women that makes them more innovative or more accommodating. Is it about our biology is there something from pardon my, my friends for something about my genitals that make us more innovative, I don't think so. I think it's really about the social niche and the social role that women played in these communities, maybe by being for example more isolated and not going to work and I'm talking about the early work from the 70s. I'm talking in this kind of more isolated women like communities rather than going to work with your briefcase and needing to be all formal and conformist. That might mean that you're more likely to in that sense makes them a bit more teenage like in the sense that they're lacking some of the conformist pressures. You can also imagine that this is about something else entirely which is about the expectation that women are more accommodating women will adopt more they will. They're more agreeable they're less rigid. And again, depending on what social that what social hypothesis you go by you can try to simulate that in the lab, some prime people with being more more accommodating or more flexible priming people with being more isolated in their group and so on and see what happens. It's not much about the factors of the specific communities but really trying to find them the reason for why and then it would hopefully be something that you can say okay this is not just a weird society pattern. And then a more or less, you know, gender equal community would expect to find a different pattern, a different role for women. But, you know, I think we're very, we're very far from really understanding this I think the first step would be to try to even just replicate these findings in the lab which I'm not sure what happened. In the lab you really strip everybody from all their roles, women and men come to the lab and maybe they don't carry their all these social biases from their culture. Yeah, I share your pessimism and trying to replicate that, but I'll wait for the paper to sort of. Thank you. You have another question hand up. Do you want to meet yourself. I have a comment on the question. Actually, the two comments one is. This is fascinating. I'm pretty impressed. And I have seen a few issues things. Another comment is that what you said about the role of women. Another role of women is, is known from the label is known under the label of the grandmother effect that women are great preservers of the language right, and it might be that it's precisely what you're saying to be two sides of the same coin. They, they stay in the familiar environment that they under under not not so much pressure to conform to the majority out there. So on the one hand they could lead their innovation but on the other hand they have the, the luxury of preserving the old forms or the old language, right. And of course I'm going to know what I'm referring to right so so it's, it would be fascinating I think to find this kind of common denominator for these two. So apparently different roles. And the question is, do you have a closed set of collaborators or are you interested in working with phonologists from the outside for example. I believe in teamwork so most of my projects as you saw there were multiple logos on the top it's because generally more brains better. We could make some assumptions about the complexity of phonological systems and also about the presence or absence of marked and unmarked unmarked are always there but the mark features right and which mark features and. Okay, so definitely definitely feel free to drop me an email when I do want to say. About what you just said, this is really important because one thing I didn't mention is women's role as teachers right so as mothers, at least again this depends on how gender unequal your community is. You might assume that children will get most of their language from their mother and not their father or from their female kindergarten teacher, rather than their male kindergarten teacher so again this is another prediction that has to do with the role of women in the society and being a being a teacher on one hand you preserve norms but also your innovations as an adult would then get transmitted as well. So it's really it's really a to a to a two way street. And thank you very much. We got another question in the chat from Deandre and I happen to know that Deandre just accepted an offer at University of Oregon to do his PhD linguistics there so congratulations Deandre. As early in the presentation you mentioned being able to describe the criteria for thresholds of being more grammatical in structure or not. We able to would you describe what those criteria were. Yeah, so this is a pretty common measure and language evolution. And all these iterated learning paradigms and we just expanded on it and adopted it so basically it's trying to find correlations between the way meanings are different and the way the strings or the words are different. So if you can imagine that words that share similar features would also share similar part words. I mean and this is in language is very hard to give this example because so I imagine that there is a green a green red, a green little chair, and a green little basket. Those things are different in one feature which is what they are and are similar on two adjectives. So you'll expect them to get the same transcription in language. So looking at the differences between what people describe and how they describe it. And in a language that's unsystematic we find very low correlations between meanings and strings. So similar meanings are going to be expressed with very different words and different words can different meanings can be expressed with very similar words. And in a systematic grammatical language you see this predictable relationship where things that are sharing features also share morphemes or share subpartwards. And when you do this, we calculate the humming distances between the meanings, and we do kind of a correlation there and then we calculate the Levenstein distances which is a measure of how words different from each other. The number of insertions deletions and replacements you have to do to make one word into another. And that shows how similar the words are, and then we correlate these two. And then we see whether indeed it is the case that similar meanings are expressed using similar words or not. And that's that's the measure. And it's, it's some math, but generally it captures really well, whether there is this grammar this part morphemes associated with meanings if a language really doesn't have it, it gets scored very, very low. Thanks. One more question from this one from Sally Coco. Yes, one initial question and a couple of comments. Are there languages that are less systematic than others. Yeah, definitely so in these. So first of all, I mean, there are, there's a lot of literature on it's typological cross diversity studies of languages today. At work, I am happy. I mean, I'm sure you may know this already, but there's some work that looked at, for example, walls and tried to correlate, you know, different features and thinking things like Lupien and Dale. They've, you know, the number of morphological effects is the number of irregularities, the number of case markings, how untransparent those markings are. And the kind of features you have that's that's considered to be more complex languages like creoles, but also English tend to have actually quite a very systematic very high kind of. I often tend to refer to this as elegant, but of course it's not necessarily more elegant, but just kind of simpler and straightforward way of grammaticalizing different features and I think these. The problem with these studies is that every study use a different metric of complexity. And when you use a different metric of complexity you can get a different results and the. I think it's all these studies tend to agree that there is differences between languages but on what and how to measure them. Then I think there's a lot of work ahead. There's a very, very deplorable bias. And the bias is that we learn languages by rules. Right. It's wrong. You wouldn't know the rules before learning the words and discovering patterns. Yeah. Second, I don't know that creoles can be learned more easily than say Japanese. Yeah. I don't know what's some of the language. Nobody has conducted that kind of analysis. Exactly. Prejudices that we shouldn't rely on. No, definitely. I agree with you 100%. I think those studies are missing and in a way trying to do this in an artificial setup is the first step trying to take a language that we have no biases towards no association and just measuring the complexity of the language seeing if it's learned more easily. Now the question is how do you do this with real world languages. And what are the who so if you take a Japanese if you try to you know to contrast Japanese English and a It also depends who is learning these languages. Some people might find the creole easier than the Japanese and vice versa so it's really important to do this study in a diverse participant pool. Nobody has done this to my knowledge. And the other thing nobody weighs something they have learned a critical mass of items and rules in a language before speaking. So I come from a. It's all incremental. For me for me that the discipline that I kind of grew up in and the way I was trained in my linguistic training is not a generative approach. It's a usage based approach relying on item based learning and statistical learning and maybe later on in life. Children can extract these kind of rule higher rules and abstractions and representations of their language but the way I perceive language learning is really an incremental step by step almost Ross learning to a certain extent until at some point you discover a pattern and a former rule, but the very beginning. Therefore that would lead you to refrain from statements such as languages that are more systematic than others because in the beginning, you know, everybody is at the same stage trying to figure out what the patterns are. Yes, yes, definitely. This is why I know you're I 100% agree with you I think in the discussion we make a very important point about the fact that the very beginning. This is indeed there is no advantage it's really later on. Some people learning a systematic language would be like, you know that they would get this realization. Hey, all these words share something in common there is a rule. That would make it easier to learn the next word to learn people who learn low structured languages. They never get this, this rule and therefore they rely more on their memory so this is the variability and learning some people are really okay with this. Some people struggle more. And in the high structure languages we just don't see this everybody forms a rule. Easy. That's a brilliant point we are not equal when it comes to learning. Yes, and when so when you deal with language diversity. There's something that cognitively should be factored in. You know if you have taught a class and you give an assignment and you see the answers. None of them is really perfect. And yet, they are imperfect in different ways. That's a question that we have to address. I, you know, we have. We vary in the ways in which we solve problems. Yeah, for sure. Indeed, and I have a feeling that all these effects that I find in the lab group size or network diversity. It's really about variability and heterogeneity at the very end it's about how much we share with other people. And that leads to the prediction. I mean this is very important of course for your work but prediction that in a very diverse heterogeneous setup where people come for example from completely different language backgrounds. The prediction there would be that they would develop languages that are very systematic. You can, some people think of this as simple but this is again not the right word to use it it's about how predictable and grammatical the language is in my, I have one project that I haven't mentioned and that is having hasn't been done yet but it's going to be done to do this group where you take people either from different language backgrounds, or you first teach them an artificial language of different types and then you have them all come together and communicate and they have to bridge these gaps and form a new language together. My prediction is that those languages would be more structured than the bigger group languages, even if it's a small community just because there's so much diversity and thing and going on, you really need to accommodate for all these individual differences. I really am very much inspired by, but by your, your thoughts and your comments and I think you're you're really right on the money. And then I'd like to start here something else with the networks in real life, our most natural, well, not natural, our most common interactions are biadic or triadic. We have the third interlocutors in the sense that we tend to interact with more or less the same people, although our diets and trials are going to overlap and vary. It's something that you can probably build in your model and see how far you go, because if you leave it where it's just a matter of random interactions with different individuals. Your predictions are going to be very different. And when we speak of emotions of norms, the bottom line is what are the norms that emerge from our biadic and triadic interactions with our preferred interlocutors. And if you do it that way, then population size takes a different kind of dimension. And I know that a smaller population is going to produce norms earlier or faster than a larger population. That's different dynamics really. No, definitely. And in the real world, it's impossible to tease these apart because small communities are typically more dense and have these kind of tight nits. And in the real world, large populations have just a different structure. So this is, I think, the benefit of these models and experiments is that only in can we look at one feature at a time. We're getting close to running out of time, but could I just follow up with a related person on that because you've sort of talked a bit about this users based background and you mentioned something about how you're showing grammar being formed through social structures and not this biological I'm wondering how directly do you see that as sort of a polemic against the Chomsky approach in the minimalist program. Was that just something you'd rather hint at and not sort of take on head on as sort of an argument against that being a necessary component of a model of syntax. Yeah, okay, big question. I have a lot of respect for Chomsky and his theories I learned studied them all very meticulously and I think they, at least at the time, were based on what we knew. And they seem very plausible and they contributed a lot to our understanding of cognitive linguistics and so on I think now at this point in time. We have not so many evidence that supports the assumptions on which Chomsky based his basis, for example, poverty of stimuli, which is, you know, a crucial kind of assumption and a principle factor in those theories today we know that the theory really is is not poor. In fact it's very rich children get exposed to you know, neural networks can learn language to very very reliable surprising degree from getting exposed to childhood speech. We know that people do correct the errors of their, their children and they do this and so there is a lot of kind of just developmental data, also showing that children are not very great at learning their language and sometimes until the age still, you know, make a lot of mistakes that we wouldn't expect them to have if they had, I don't know, an innate capacities or a principle and parameters, a component. So I think this is what good science is about. It's about coming up with an engaging theory that pushes the field forward, creates, you know, a pressure to make more experimental observations and test this theory. And then some parts of the theory gets refuted and replaced and other parts maybe get preserved, but that is what science is about. So I'm not, I don't think it's about being against it's about making it, you know, developing the theory more and more based on data. And I'm very happy to be convinced that I'm wrong. If I get presented with data from child learning, showing an innate bias, very happy to change my views and some really about kind of being data anchored. And I think, sadly, there is not so much support for currently for the Chomsky and approach. When you look at child learning patterns. I want to be a bit respectful of everyone's time. So I know there's some more questions and more questions in the chat too. I think what I'll do here is say thank you and stop the recording but if people want to keep chatting, we'll either pass on the Moore's email address or if people are able to stay in chat that's fine too. So let's say thank you to Lee more for this presentation and for engaging with everyone's questions. Lots of interesting topics, of course, are coming up and we could definitely continue to discuss this very long but thank you so much for joining us. Thank you. Thank you.