I'm going to turn it over to you and hit the record button. Sarah
I'm Sarah Brown-Schmidt from Indigo
Daphna Heller from the University of Toronto. And we're gonna start by giving you a little bit of history of Common ground Common ground is the knowledge that is shared between interlocutors and is mutually accepted as such. So it's not enough that it is actually shared, but it has to be acknowledged as shared. And this notion is over 50 years old that comes from the philosophy of language, philosophers of language notice that language in conversation, the Felicity of language and conversation depends on the knowledge that interlocutors already share. So, they posited a set of a notion of a common ground. And in some like returns, this is the set of mutually accepted prepositions. So this is information. Now, Herb Clark has brought this notion to psychology. And he was concerned not just about what this information is, but also how it becomes shared, noting that this is inferred from indirect sources, and he noted that shared information can be visual, or it can come from language, or it can come from background assumptions about the interlocutors. And he has pioneered, what we know is the tangram toss a version of the referential communication tasks were too naive partners talk about a director gives their address see instructions on how to reorder those nameless, originally nameless have figures. And two observations come out of those situations. One is that over time, the referring expressions that speakers use become shorter. And the other one is that there's a movement from indefinite descriptions to definite description then that has drip shedding attributed to the accumulation of common ground. So in terms of the influence of these early observations and theoretical proposals, in terms of the influence on the psycholinguistics, empirical literature, there's been four kind of prominent views the earliest came from Boaz, qaisar, and his colleagues, who posited an egocentric first model the idea that Common ground influences language processing only at a second stage or after some sort of delay. One prominent response to this came from the constraint based theoretical approach, arguing that common ground is one of many probabilistic constraints that influences language and this was prominently argued by a paper by hand on my Canon hillson colleagues. More recently, Courtney and Garrett posited that ordinary memory account arguing that basic memory processes could give rise to phenomena that look like sensitivity to common ground, without necessitating representation representing Common ground as such. And similarly, the interactive Alignment account posited by Pickering, Garrod proposes a mechanism by which through basic priming mechanisms, conversational partners could come to have similar representations of language to use similar terms, again, without representing Common ground as such. So across all of these different theoretical proposals and empirical approaches in the psycholinguistics literature, one observation that non I have made is that there seems to be despite various differences about exactly how this works at consensus, for there's some need for representations of shared or Mutual knowledge is the basis for communication. Another observation that we've made is that, after about 40 years of empirical research in psycholinguistics, to dominate dominant types of questions seem to emerge. One we've described as a cataloguing approach, where there's a large number of papers now documenting the different sorts of parameters or situations that modulate when Common ground will or will not influence language processing. Likewise, there's a large number of empirical papers examining timecourse questions so when as you're producing or understanding a sentence is common ground to language processing. And here, what we're going to try to argue for is a different perspective, we want to shift the focus of empirical psycholinguistics work on the role of perspective and language to new questions.
So over the next couple of slides, we are going to try to convince you that a focus on Mutual knowledge or any other kind of shared representations has limited utility. If we want to account for or model the wide range of conversational phenomena that we see, we have to widen what we're looking at, we can't just look at what is usual or shared, we also have to look at the individual perspectives of the conversational partners. And our first example comes from Questions. And we are going to demonstrate this first with the case of the pig with the hat. So imagine a scenario where Sarah on the right and me on the left are talking about are looking at this display or looking at this displays from two different perspectives. And what you would notice is that some of the animals here have a white background, that means both of us can see them, some of the animals are completely blocked from either my view or Sarah's view, and the corresponding animal has a gray background. So each of us also knows which ones the other one can see. And what Sarah and colleagues has shown in a series of studies is that what speakers ask questions about information that they themselves do not have, but they assume the partner house. So here the question is what's below the pig with the hat? Now you've located the pig with the hat. And you see the Sarah's asking about my private information, we also want to point out is that questions are never the less sensitive to mutual information. Because notice that here, there are two peas, figs. And we have this in the question says pig with the hat. So we have to specify which big this is. And this is our first hint that speakers and listeners need to be sensitive to both private and shared information at the same time. The second case, we'll consider we described the case of the knowledgeable dentist. So imagine, I have taken my son James to the dentist, we've got James on the left and my dentist on the right. And I start panicking and think, oh, did James brush his teeth this morning? I'm gonna be really embarrassed that he didn't. And so the question here is also about questions. But here, the theoretical question we're asking is, who do you ask? So if you need to know this information? Did James brush his teeth? Who do I ask that question to? How do I seek out the right person to ask? Now, if I was simply going to choose which person to talk to ask this question to, based on say the quantity or proportion of associations having to do with teeth? Clearly, that person would be the dentist because all of my conversations with a dentist has been about teeth. And and in fact, probably most of the conversations that I do have about teeth are with my dentist. On the other hand, most of the conversations I have with my son are about other things like don't hit your brother. And so in terms of pure Association account, you should ask the dentist, but definitely I have this strong intuition that if you need to know whether your child brushed his teeth this morning, you asked your child why because he is in the best position to have direct information about whether or not he actually did so. He's the one that would refresh to see if they're not. Whereas the dentist could appear inside his mouth and look for plaque. That information would only be indirect. So our intuition here is that in order to decide who you're going to ask a question to, you have to model the likely knowledge seats of different candidate people, and also the sources of those knowledges. Now maybe you're skeptical, let's change the scenario of it. Let's say Dan Greiner came to visit. He's a great chap, and I'm really busy. I've got to go prepare for this webcast. So it's not so I see Dan, could you please take my kids to the dentist, do this. Okay. Now, Dan has never met a dentist before. He has never talked to my child about tooth brushing before. But yes, Dean can still model the likely mental states of both the dentist and the Son and know that the dentist is only in a position to have indirect evidence as to whether James fresh in the seat by curing himself and looking for plaque. Whereas James is in the best position to have direct knowledge about that fact, whether or not he did so. So again, our intuition is here is that Dan in this scenario would ask James critically, the case of ineligible dentists illustrates in order to do this very basic conversation will move, you need to model likely mental states of other people that have nothing to do with your mental states that are different from your own mental states.
Now, maybe you're still skeptical, and you're saying we knew questions were about about private information. And not a lot of people have looked into that in in psycholinguistics, perhaps. So Common ground, or shandra presentations is going to do a very good job in explaining reference. And to maybe put some doubts in your head about this, we are presenting the case of the Deceitful Lego. So these are cases where in the display, this is a referential Communication task, like the ones that have been used before. And in this display, there is one object that's a normal object except the crayon. And then there's a second object. And all of you are looking at this second object and thinking, Oh, this is a Lego, because this is how we normally infer functions of objects by their appearance, except this is a Deceitful Lego. So it's not actually it doesn't actually have a function of Lego, it has a function of the crayon. So really, we're dealing with a deceitful object. Now, crucially, I know that this is Deceitful Lego. I know this is really a crayon, whereas Sarah does not how do I know that because, well, I tested it before. What's important about this situation is that the function of this object cannot be specified in Common ground. My perspective says this is a crayon. The function of this object is a crayon, where a service perspective says the function of this object is a Lego. So we do not have any piece of information to put in Common ground for the function of this object. And crucially, if you look at language production patterns, or specifically, the production of referring expressions, we don't find that people, speakers produce referring expressions that are sensitive to their own perspectives, we don't find that listed as speakers produce referring expressions that are sensitive to their addressees perspective. But rather, we find an intermediate combination of both. The next piece of evidence we'll consider comes from the literature on memory for conversation. And this was a literature that goes back to the late 70s. And it's pretty consistent, and that the summary that we can take away from that literature comes from studies where they have people come together to have a conversation. And then after delays have anywhere from about five minutes to three days, they say surprise, recall everything that was said, or in other cases, they use recognition memory procedures. And there's two central findings that are relevant to the present situation to present arguments. One is that after a short delay, so five minutes to three days, people can recall from anywhere from about five to 40%, of what was said, which is quite low. The second piece of information, and perhaps the more important one is that there is a strong, egocentric bias. So there's about a two to one ratio, approximately of what you can remember such that people are much more likely to recall and successfully recognize things they've said themselves, compared to what was said to them. And so here, we want to ask, what are the implications of this asymmetric representation of the discourse history for language use? Okay, and so, here, what we'd like to argue is that on the basis of these memory findings, it would suggest that distinct representations of the discourse history are the basic basis of conversation. So here in the circles, I'm representing, on the left of those memory for the discourse history, and in Brandon, illustrating that it's biased in terms of her own contribution. So mostly, she's going to remember what she said, she'll remember a little bit of what I said, But mostly, she'll remember what she said. And I'm similarly going to remember most of what I said, and very little of what she said to me. And so if we imagine what is even possible in terms of shared representations of what was said, it's a very small piece of the pie. So while it's been long known that these representations of conversation are ecosummit egocentric and asymmetric. Here, what we want to do is pivot and ask, okay, well, what are the implications of these distinct and asymmetric representations for language use? Now will point out that ever since the idea of common ground and has been introduced in the literature by people like Clark and Sal Necker, there has been this idea that Common ground is represented distinctly in each person's mind. Here, we want to say, Okay, well, what does that mean for a theory of language use and the role of perspective in language.
And so what we're going to do today is outline our new proposal, we termed it then multiple perspectives theory, and it has four components. And we'll walk through each of them includes a representation of oneself. So what do I know a representation of the other person, a comparison process, and then a short lived representation of the surface structure of what was said. In my little animation comes to demonstrate that those two are presentation of the self and the other are going to be separate or presentations. So what is the presentation of self? What does it include? Well, for me, it includes that I live in Toronto includes a lot of information about the Toronto restaurant scene, it includes the all my kids hockey practice times, and it also includes some information that there was a place called Nashville, I don't actually know very much about that place. Now let's consider what I'm assuming that Sarah knows. And it's important in this context to remember that I'm in theory inferring what Sarah knows from indirect cues. And as such, because it's inferred from indirect cues, this representation is probabilistic. So first, there are bits and pieces of information that I assume, I assume, Sarah to know, we call these the full color representations. I know that I assumed that Sarah knows that she lives in Nashville. And I assume that Sara knows that she works at Vanderbilt and I assume that Sarah knows that she drives a truck. I also assume that Sara knows things about the national restaurant scene but I don't actually know what these are. I also assume that Sara knows the hot the she doesn't know hockey that they don't play hockey and national. I Sorry, sorry. prades. I know but ever since PK Soobin last I feel a little Um, so the soccer practice time, I assume Sarah knows them. But I actually don't know what they are. And these are we call an outline or Placeholder representations. So these are things that I could ask Sarah question about, for example, if I if I want to find out and notice that that's not does not have on their representations of the island I grew up on because she doesn't know where I grew up. So she just asked me about that particular Island because she's ever heard about it. Okay, so we've got self representations and we got representations that the other The next component of the theory is a comparison process that compares the two and for now we're going to say that this comparison process occurs a conversational turns and compares these representations of self and other resulting an output that indicates the similarities in these representations as well as the differences and this representational output is used to update representations of self and other and so we're going to use the case of the misleading words to illustrate this process. Okay, so case of the misleading Martini imagine the man and I are at some sort of party and this is timeline and on the left we have the first beliefs in yellow so the if not believes about that she wants another Martini does not also believes about Sarah, that I know whether I'm in a position to one another Martini this so because Stefano knows that I know whether or not I want another Martini that's not says to me, Hey, you want another Martini? I want to Sarah. Okay, so I know that I'm pregnant and I'm drinking a lot. And what do I know about the nylon? She's Canadian, she probably likes hockey. So those are the things I know about if not and myself. So but then if not asks me Do I want another Martini? This prompts this comparison process where I know I'm drinking a mocktail and now this causes me to realize does not think I'm drinking a martini through the comparison process this allows me to conclude itself not just not know that I'm pregnant. Which brings us to time three so my notes the undertaking a mocktail does not asking the questions whether want another Martini causes me to realize that if not things and drinking another Martini and therefore doesn't want pregnant. This allows me to say to death not I'm pregnant. You're pregnant.
So in order to account for echo questions like this one and other things, allistic phenomena that are dependent on the most recent discourse history, the fourth component of our theory, the fourth component of our cognitive architecture, which we fondly term, the reverse teleprompter is a surface linguistic representation of the most recent discourse history. And the way we think about this, this is why we call it the reverse teleprompters. In a normal teleprompter, there's text and the person who's reading from the teleprompter is saying the text here instead, people are conversing, and whatever they said goes on their own personal reverse teleprompter. Now, this would account for things like echo questions, contrast and accenting different kinds of Anna fours, syntactic persistence effects, and probably a lot of other things. And what we want to point out is that this is really not a common ground like for presentation. However, because of the recency of these distinct individual memory representations, we are in a very good place as conversational partners to assume that we have the same representation, or at least very similar representation on the reverse teleprompter. And so this representation could superficially account for phenomenon that have been previously attributed to a representation of common ground. So this proposal leads to a number of predictions which will describe now the first one is, has to do with a language. And what's important here is that we are bringing assymmetries into the picture. So our proposal because of the nature of having distinct representations, is and and will do different outputs of the comparison process is predicting that language is going to be sensitive to distinct representations. So we went out in the world and looked for places where linguistic markers are sensitive to distinct representation, and one such place has to do with evidential markers. So first, let me introduce you very, very quickly to evidential markers. 25% of the world's languages, grammatically mark the source of evidence on Assertions or Questions. And what's important for our purposes here is that some of those evidential markers are sensitive to a symmetry in perspective. So for example, in Korean, and in Japanese, there's a direct perception marker that let me focus on Korean and Japanese a little more complicated. And of course, the person who's the expert here on evidential is is Elgin me. In in, in Korean, when I'm looking out the window, I can say it's raining direct perception, only if my addressee is not experiencing the same perception. So if I'm talking to someone on the phone, I can do that. But if we're both standing and looking out the window, I cannot do that. And there that is a marker of a symmetry.
The next prediction comes is that by truck byproduct of these asymmetries and memory that we pointed to, so imagine that if not and I are wrapping up a meeting, and I say, oh, okay, I got to leave a little early today. That's not I've got to go to this workout with my favorite peloton. Instructor Christine, and this is all what what kind of music does she plan? I'm like, oh, stuff from the 80s I love her pet that guitar rides. And I said, Oh, that sounds like fun. And then we, you know, we wrap we keep talking a little bit about, you know, perspective asymmetries and so on and then say, Okay, now I got to go to my 80s, right. And in the bottom right corner, you'll see a photo from the four times of me Actually, with Christine at the mothership having a lovely life writing class, but any of it so I say I gotta go to my 80s. Right. So what's interesting, from the perspective of the fact that we're gonna have these asymmetric representations of the discourse history, for me, at least right, it's a well supported definite, definite description. It's this right I'm going to I totally know what I'm talking about. And I'm likely to over assume that Noah's ability to unfeasibly understand this expression, this same expression, which is super easy for me to understand or produce is likely to require some action Combination from her perspective, she'll be able to work out what i'm talking to. But because her memory of this discourse that we had is different than mine, she mostly remembers what she said, pay attention to what I said, She's, she's gonna have to work harder to understand what I mean by this. And so the theory predicts that in conversation, people are likely to produce these definite expressions that from their perspective are quite easy to understand and produce. And because it's based on their essentially bias representations, that is force history, and in the listener, while they may be able to accommodate them, they're going to be much harder to understand because they have a different representation of the very same experience. So this brings us to some conclusions. And what we've tried to argue here today is that representations of what is shared have limited utility and accounting for the wide varieties of language use and the different kinds of conversations one can make. focusing exclusively or primarily on shared or Mutual knowledge is not going to provide a very good account of questions in terms of who you ask, and what do you ask them, it's not going to provide a very good account of communicating in the absence of common ground. So cases like the Lego, and it's not going to provide a very good account of cases where you track and Mark sources of evidence and differences in sources of evidence across. And our aim here at the Multiple-Perspectives Theory is to explain existing phenomena with this new cognitive architecture to generate novel predictions that we can go test out and perhaps falsify parts of this theory are, you know, affirm parts of the theory, and the shift the empirical focus to do and different kinds of questions sort of beyond how definite referring expressions are understood and conversational against other kinds of questions having to do with things like informational questions, or different sorts of grammatical markers and composition, things like that. Thank you all for listening. Thank you to our institutions and our funding sources, and the many people who've had lots of conversations about this topic with us.
All right, thanks so much for the lovely talk.
I am going to, I'm going to turn it over to Erce and Vic to facilitate the discussion and questions. But before I do, I want to
say a part
of this is going to involve questions from the audience. If you've got a question that you'd like to ask, I'd like you to go ahead and put it in the chat, send it to me. And when you do, please note your career status. So what we're going to try to do is make sure we've got a good mix of questions from, from people at multiple different career stages. So just send them to me. And we'll keep a log of them. And we'll kind of work them in throughout the discussion. And so without further ado, I'm going to turn it over to Arjun Vic.
Hi, Sarah. Great talk. So I'm gonna have a question about something about the memory representations about these conversations. So you suggested that, you know, we track our own perspective and the others perspective and compare them and, you know, update our representations based on the outcome of this comparison. And this sort of feeds into, you know, like, the linguistic choices we make, and so on. But presumably, this also has to feed into our longer term memory in some sense, right? Because we don't keep saying the same things over and over again. But yet, you also mentioned that memories are quite fragile. And in fact, people are quite bad at actually tracking the source of their memories, right? They not only remember more of what they contributed to a conversation, but sometimes they may also act as information they learn indirectly is actually something that they directly experience. So I guess given these asymmetries in the long term, like what is encoded in terms of like the mental representations of the self and others in memory, and what are we more likely to remember later
on? That's not my PT answer this question. So yeah, so the question is like after some longer delays, what all is remembered,
and
probably the estimates that we get at short delays are a good indicator, the likely patterns will see it longer delays, and forgetting will occur over time. That said, the I have not like my knowledge, seen anything about say, like the asymmetry that you see in memory for what you said versus what the other person said, is present. After very, like minutes, and it's also present after days. I have not seen anybody tried to track the time course of that asymmetry over time. But I think that there's a potential for pretty interesting interactions between these memory asymmetries, and then what people are likely to talk about in terms of perhaps even compounding those differences over time. So Dominic Knutson, and forget his first name, Libby go is the second author of a series of papers looking at what people are likely to talk about in conversation. And there's also an egocentric bias there. So people are likely to keep repeating what they said before. And so it's interesting, you could imagine that in addition to a sort of this asymmetric initial encoding, you could just keep repeating yourself. And that's what I like, magnify that over time, and but fleshing out those critical questions is remains to be done.
Cool. Oh, I'll jump in with a question. And for those of us for the for everybody in the crowd, where we hold during your questions, and as well, and we're doing this as interleaved as possible. So as you have questions, please send Brian Brian's gonna send them to us. I noticed one person raised a hand, I think that would be good at that question. I think went through Brian's I think that's how it all be immediately. But we're not doing a thing where like, or Joe and I asked a bunch of questions, and then everybody asked their questions, and we're trying to make them as as mixed up as possible. Okay, with that. Um, so, definitely, and and, Sarah, you talked a lot about some of both keystores results, one of the things more fun results, maybe is that he's done this stuff, where he's looked at people's estimates of shared knowledge as a function of their familiarity with those people. And he sets it up in a nice sort of dramatic contrast of saying, you know, someone really, really well like your, your spouse or your partner. Does that mean, you're really good at distinguishing their information that they really have versus not? Or does that mean, you're actually really bad at it, because you assume you share so much that you actually share all your knowledge. And the bottom line ends up being it's the ladder that that we we seem to assume that people we know really, really well have more shared knowledge with us with, with ourselves and others. Maybe you've experienced this phenomenologically, where you've gotten annoyed at your partner, because you assume they must know something that you know, and then you think about it for a second, you realize how could you possibly have known that that would make any sense? I just assumed they did. Does that I have a feeling that that fits reasonably well with this framework? And if so, how might that be?
I'm Vic I'm familiar with I think I've seen that result in like maybe three different papers, I think, did maybe even as a result, like I see. But my take on that was always that this sort of over assuming for the very familiar partner. And being counterproductive was in part because the target domain is novel to both. And so if we're if Dwayne and I, my husband, Duane and I are playing a Tang grande game, where we both know nothing about the figures, this over assuming as is likely to be a problem. But if instead it was a very familiar game, like when his soccer practice, over assuming his knowledge, maybe would be less of a problem. And as
a great difference, because you guys know that you actually do know that stuff so much that the likelihood of overestimating is lower is. So
if I might my I don't know that this has been explored in that literature or not. But my my prediction would be if it's a target domain for which there is actually a lot of shared knowledge, that the over assuming would be less problematic, or at least more or less inaccurate. But in terms of the relationship between the degree of egocentrism with a partner and and how that would play out in terms of our theoretical proposal, that's not if you want to, yeah, so I think it has to do with something that has not perhaps, been given much attention, which is those probabilistic cues that you use in order to represent what the other person knows. And I think the more I'm more familiar you are with a person the more interactions you have with a person, you are going to have more cues into what they do or do not know. And so there are more avenues for you playing true for those cues incorrectly because remember, these, this is not really about what we extensively share or not. This is about what we assume the other person knows based on indirect probabilistic cues. So I think that it I mean, it's an empirical question, but it seems to me that the more cues you have to work from, the more likely you are going to be to draw the wrong inferences. That could potentially account for why you're going to make more mistakes with someone that you are more familiar with, where you're supposed to, like, you would think it would go the other way. Yeah, you're supposed to maybe those cues are I mean, intuitively, you think I really know more about what they know. But are these cues really working in in the in the right ways, if you will? Right. I think that's probably something that and those audience design papers are also sort of reinforced by early work by hustling Krauss where they had people estimate, you know, what are examples for these hilarious movie stars from the 70s. So I only know a couple of them. And they're, you know, people basically overestimated the degree to which people would under know, these movie stars if they knew them themselves. And then more recently, scot free and dark. And Jonathan, tell us how they're really interesting line of work looking at judgments of learning, but you were you're judging other people's learning. And again, they observe these egocentric type biases. So things that were easy for them to learn, they were estimating that others would also find that easy to learn. So I suspect that these biases are likely to infiltrate a lot of these sorts of prospective comparison and representation processes in a variety of domains.
Cool, or just Should we go to an audience question? Yeah.
Maybe we can start with a student question. Okay. algorithm. Okay. So, this is from Chiba, Japan, I hope I pronounce it correctly. And she is a PhD student. And she has a question about the type of information sharing so well, more salient information from the speaker, he remembered more by the listener? And is there a way to use this to increase what it means and maintained in the Common ground?
Sorry, I feel bad, because, um, I know jido, that means you should answer this question. I don't, Hi, nice to meet you. Um, um, there are a couple of results that I know about the top of my head indicating that your intuition is correct, that there are going to be predictors of what's likely to be remembered. So one example comes from a study that was motivated by the Anita Hill hearings, and where she testified that then Justice Thomas made some inappropriate comments to her in the workplace. And then the experimental result that goes along with this tested people's memory for inappropriate comments as a function of whether it was a bar or workplace setting. And they found that the inappropriate comments were much more likely to be remembered, particularly if they were in a inappropriate context, like an office setting. And so those findings, and then some other related findings, we probably can't do this kind of work anymore. They kept they tested people's memory for Brian, I saw I can't remember the citation right now, but it's some a paper where they tested people's memory for comments made in a seminar and the professor's new photos. Anyways, the jokes were well remembered. So there is evidence that salient, inappropriate, contextually unlikely comments are likely to be better remembered by fault. And so that's the case, I think, where you might get more symmetry.
There was a loose bass result of the question, the numerous comments won't move on. Dude, I hope that that was a satisfactory answer to your question. If not, feel free to call them task. All right, I will call on so for those who are comfortable asking a question themselves, we're just going to ask that person to themselves ask a question. So I'm Clayton Rockwell, if you wouldn't mind unmuting, and asking the question.
Yeah, thanks for taking my question. So I'm a PhD scientists in government. And I was interested in the reverse teleprompter notion.
So my read of her Clarke on the recent dialogue history is that there's evidence that sometimes we are implicit, but that the recent tile on history has entered the common ground. And I understood your main point is that Shouldn't reverse teleprompter shouldn't be considered in the Common ground, if you will. And so I wanted to clarify if I understood that correctly, and then maybe ask you to expand a bit on why you would consider that in progress.
So for us, one, so we realize that we're pushing for a strong position and saying, we cannot, there's there's no real way that we have come up with maybe someone else will, to show that if you have also ever presentation of common ground, you won't be able to navigate a conversation. Instead, we're saying, okay, you're going to need to have your own representation of self and the representation of other and you're going to need the comparison process for those things that you did not previously have a representation of other four. So these are some differences. If you had a representation of common ground, maybe the similarities will go there. But maybe there was no, you from before that becomes activated during the conversation for you to have that. Now. Now, to your question about the reverse teleprompter. This is not meant to be a Common ground life presentation. This is supposed to be a surface representation that you have where you maintain your own linguistic representation. And it's only because of recency that people are able to assume that that I can assume right now that your reverse teleprompter looks like mine. So I can, um, maybe I didn't do that once. I don't remember the you see it, like your question has already fallen off my reverse teleprompter. And so I cannot pick up any pieces of it and demonstrate that with things like VPN ellipses or anything like that. That helpful. Okay,
so that it's really the focus on the surface form such that maybe the semantic content of my question and illocutionary force in my question is present. But the words that I use to express it are gone. And that's the main idea. So
I think it's an it's an empirical question, what level of linguistic representation is going to be available in the, in the reverse teleprompter? Do we need to? Are we just representing words and syntactic structures? Are we like, Is there anything else there? I'm kind of hesitating to give example of what anything else could be there? Because I really don't know. Right? I mean, we could not, in in one paper do justice to the last 50 years of empirical research. So so we don't know this is a this is something that we're hoping to look at in in in further research. but crucially, it is not something that we are assuming to be shared. And it's not something that we are assuming to be long term. Sara, like some early evidence, actually, when I first gives her first papers on this topic had a lovely demonstration of how like, recently mystic history that's not Common ground can infiltrate these prophecies that were they like played words and a headphone and I don't remember the exact result. But it was basically influence the language use in the moment, even though it was only heard by one person in the headphones. And the general motivation for the short lived representation of a very recent discourse history is also inspired by garrison psykers proposal about the structure of discourse as well as some of the Marilyn Walker's early writings on looking at how pronouns link up with their preference.
So just as some fare ID, maybe we can hear from Craig Roberts. Hi, thank you very much. I really enjoyed your talk. And I think he's really stimulating questions you're raising. So but the basic question I have some background to that is, it sounds like you have some very good arguments that common ground is not sufficient to capture the kinds of interactional constraints you're interested in, right? But it still seems like common ground might be taken. To be necessary, for example, for explaining felicitous uses of pronouns versus definite descriptions, or expressions like to proscenium, xotic, focus, things like that, they miss something that are used that this is not the case that in other words that you don't need, it's not sufficient, you will need a richer representation of the knowledge base of the participants. But does that mean we don't need common ground or something like Congress anyway. So we, um, so I will address the specific linguistic phenomena you've just raised, like, like to Oh, or like, DPI ellipses, or like, for xotic, trapping, accenting or anything like that. And we went once. So I think it's very, it's very hard to actually look in the literature for actual evidence for that. But it seems to me that these are phenomenon that are licensed by the most recent or the one utterance or the one that just precedes them. And those actually are not licensed by prior discourse history. So that makes us hope, or think, or try out to see whether we could do rise these surface linguistic licensing in the reverse teleprompter. So it's not about having a shared or presentation. But it's about the fact that these are reasons that you and I happen to have the same representation on our reverse teleprompter. And that's what licenses those phenomenon. First of all, I actually want I'm not going to argue that here today, but I would argue that the recency is way over claimed, I mean, maybe the majority of the market that's not the case for premiums, for example, or even to to possibly so we could argue that I mean, if you look at cross assignment, we decided, you can look at the fact that they have these examples where the intentional structure of the discourse could make something that was only mentioned eight or 10 turns previously be the salient FCV for a hearing it so one, but second thing respond to that. Okay. So I would say remember that the information that comes off of the reverse teleprompter doesn't go in the garbage, right. Um, it is used to update your, the representation of self and other in a non linguistic fashion. So, those representation are not meant to be representations of language. They're meant to be representations of information. And if you ask me what that means, I will say, I don't know. But I don't think anybody else does. This is not recently have mentioned return something. So we have to be careful about that. That's my
point. Right?
Then the second point is, I mean, it's important to be careful not just to look at cases I understand, like, you might just say, Oh, it's one thing to say in the abstract that you need this notion of common ground in order to get at certain presuppositional phenomena or other phenomena as well. But it's another thing to say that that's actually how we represent common ground. So we have this thing called the common ground representation. Instead, I know, I know that there's a spotted pig with a hat on and I know, from the
representation, I
have that, you know, we can call that common ground. Okay. But it doesn't mean I have something over here called the common ground on on joint representation, I think it becomes even more interesting. And I'm just going to leave it at this and have the conversations more even more interesting when you have like large groups of people who are talking and there it becomes important and very difficult to track individual representations for a large group of people. But something about the common ground of the shared interaction then does need to be represented if we're going to capture what the speaker feels. They are they can solicit as they say, so I just I just want to drop that into the consideration that there might be However, it's represented this abstract notion might still be notion necessary, if not, if not sufficient, I mean that I have no question. So make a quick comment, but then Sarah is going to talk about the larger groups. So, my quick comment is that as as in the in the psycho linguist, we are concerned about cognitive architecture, so So that's what we are interested in. And what we're arguing is that in our cognitive architecture, Common ground has limited utility. We are not arguing that the whatever is Mutual knowledge is not going to play a role in language. Yes. So that's a really important distinction that you're making. And, Sarah, do you want to talk about groups? Yeah, if I'm the so the scene perception literature, the visual perception literature has some really interesting concepts that court and I think, an interesting meta way to that problem of how you represent knowledge in groups. And this was the focus of Shannon's dissertation and some ongoing work in her lab at Iowa. Danielle, how people represent the knowledge of groups. And so one of the ideas from the scene perception literature, so people can extract this summary representation as well as representations of outliers. And so she has been exploring this, this analogy in the knowledge representation domain asking to what extent can people represent the average knowledge of the group on a particular topic? And then, you know, wherever she takes her from there, so
long conversation, we can have a lot more fun. Thank you. Thank you. So continuing with questions from the crowd, Tatiana? Sure. Do you want to unmute your question? mute?
Can you hear me? Yes, good. Okay. Thank you for a really interesting talk. I particularly also appreciated the dynamic to and fro between the two of you, I thought that that was a really excellent prospect. Um, I'm trying to think of, and you probably been thinking about this for a long time. Concrete variables are ways to measure to test your predictions that in fact, you don't have this separate core of common ground, that's realized, linguistically. And so my question is that I was a little bit to see what my question is, um, I was confused about the use of that one way you could test your proposal was to look at the this issue about memory, and I, I didn't quite follow, like how you would test it. So you could talk to that, oh, there's this example where people don't really remember what the other person said. And this goes away very fast. So if you could explain a little bit better about how you might test that prediction that that's one of those domains that you were looking at? Thanks. Um, I guess in terms of an empirical approach, that if it's true, that things that you produce your remember better, and that has implications for processing, what it would suggest is if there was some topic or reference that I had mentioned, but but not had not, compared to a case where both of us had mentioned that reference,
and that,
if I later refer to in the example we use, it was my ride, that if we measured the nose comprehension of that expression, it would be worse in the case where she had not actually uttered content or related content in the past. So that would be one empirical angle towards investigating whether production driven Memorial assymmetries effects language, subsequent language use. Because notice that specifically, the example that we chose my 80s rides, is such that it is very hard to accommodate out of the blue. So if you say my dog, if I say my dog, and you didn't know I had a dog before, then you can figure out I have a dog, and we can continue with our conversation. And this is something that has received a lot of attention see in the semantics pragmatics literature. So here in this ad drive, it's really hard to accommodate. And that's why we're, we would tailor this example to illustrate its dependence on a representation of a prior conversation. I have no idea how to answer this question. Oh, yes. Sorry. Sorry, I couldn't unmute. So just to so this would be in a conversation you would measure the number of times that an interlocutor brought up a topic that was novel in response to a previous topic. That's what I was trying to like this idea of the 80s comments. And then the other person didn't understand that comment. And you would measure the degree that they don't understand it. And you measure that degree that they don't understand it by measuring like a direct query about it, perhaps or like, just that little bit positive windows do sorry, Sarah. Oh, I was just gonna say, but nine, I have a lot of practice coming up with clever ways of measuring conversational understanding and interactive settings. In the past, it's been whether you look at the turtle wearing glasses below the pig with the hat and how quickly and but I mean, I'm sure there will be other methods of doing so. Okay,
thank you.
So moving on, maybe we can take the question from Josh, we Rick was a PhD student.
I have thanks for taking my question of really wonderful, interesting talk. So early on, I think you mentioned that accounts of Common ground have have differed as far as whether Common ground is represented explicitly, or whether the effects that we attribute to Common ground are emergent from memory or other cognitive devices. So could you talk a bit about how you view your account regarding the explicit versus emergent question? And so I, okay,
so I think that's, we do, I mean, okay, the fact of the matter is that we do not have an explicit representation of common ground as part of our cognitive architecture. So if this is what defines emergent, then we are in the emergent camp? Um, it is. Yeah, I think that's the shortest answer to this question. on whether it's, um, there seems to be a correlation in the literature. And maybe that's why you're asking this question. There seems to be a correlation in the literature between the rubber representational
between
in the time course questions, there seem to be a correlation between people who posit an existing representation of Common ground that is recruited in real time processing, and people who say that this information does not affect early processing. And it's delayed and some or it's non existent, or however you think about this. And these are more likely to fall in the emergence camp. But I think that that is for lack of a better term, possibly just a misconception maybe. I mean, I don't know that you suggested that I suggested that. But maybe this correlation doesn't really exist. I'm looking at Sid Horton, right underneath my own video, and I'm like, yeah, this this correlation doesn't really exist. I said. Was that an answer to your question?
Oh, yo, yeah, it was Yes. Okay. Thank you for interpreting my question in a way that was much better than I asked. I really appreciate that.
Okay, so we're going to turn to another student question. A really clever one, from someone with a Turkish name. So I'm going to defer the naming the naming of the students to my colleague or Joe from bohart.
Can you ask your question, but Hi, correction?
I probably couldn't pronounce that one.
I was wondering if so this means that paraphrasing in conversation would make it more likely for us to remember, but the other person by making it more likely to create a common ground on what you remember of that conversation. And make maybe, of course, paraphrasing often can disrupt the nature for conversation, but I can use onions for the things that we want to remember, especially be helpful. Yeah, your intuition is exactly right in the memory literature, they call this retrieval practice effect. So if you study or hear information, and then practice retrieving it from memory, you're far more likely to remember it later. However, there are some really interesting companion phenomena that occur when you do retrieval practice. And this generally comes from the list memory literature. However, there are some interesting results on social sharing. And this I'm thinking of is alene Coleman for he's at Princeton. And so and that had to do with a city where people talked about their love and experiences, is there these retrieval practice effects where if you, if you like, hear somebody say something, and paraphrase only a part of it, it can actually hurt memories that things you didn't practice. And so there could be this interesting interplay such that if whatever you practice will be boosted in memory that the stuff that you don't have to should actually be suppressed somehow, as a result of that practice.
Do you imagine something further as well, which I took from the printed version of this question is that she said that what you do, though, when you paraphrase that there is a little bit of a self bias and how you end up paraphrasing it? Is it going to cause a little pernicious effect of self contributions to being bigger than the other person's? Okay. Are? up? Okay, I think, urge I think the floor
Yeah, sure. Um, so maybe we can continue with the question from Baron. Because
they're in bed longer with us. Okay. Yeah, I'll pick another one.
So, um, do we have famous students?
That's me.
So thanks for this really cool talk.
It gave me actually a ton of food for thought about how this connects to stuff in the semantics pragmatics literature that focuses on individual discourse commitments instead of common ground, which also goes back to 70s philosophy of language. It's a less market share for most of the history of the field, but is increasingly dominant, actually, in the semantics, pragmatics literature. There's like, a bunch of things that I would love to talk to you about, but I'm going to pick one now. And that's the this literature both in terms of common ground and in terms of individual commitments, in the semantics, pragmatics literature has focused very carefully on public representations of beliefs. And that's because you know, you can do things like lie or tell completely made up stories that we all know were made up, sorry. It's just to amuse ourselves. And so these public representations, maybe in certain cooperative contexts, but not necessarily reflect what people actually believe. And it seems to me like your focus on, you know, what is an individual's probabilistic representation of their interlocutors beliefs might actually interact with this in a really cool way. So I wanted to like first observe that if I have some representation of your beliefs, and then you assert something that's not compatible with my representation of your beliefs, I could do in principle, at least two things. I could change my representation of your beliefs, assuming that you're sincere, to come in line with what you said, or I could remain committed to what I thought you believed and take your assertion to be insincere. And I wondered if there was anything in your model that gave us something to hold on to in terms of understanding when people might take one strategy or take the other?
This is a really
cool points that we totally haven't thought about. I think your question to me translates into what is the relative probabilistic weighing all different types of cues. So if I am talking to my partner And he says something that I have a lot of cues to know is false, then probably, the single utterance is not going to change my representation. But if I meet you for the first time, and you tell me that you live in New York, then I am going to assume that you live in New York, because that's the, that's, that's sort of all I have to go on. And notice that this, this was a very cheeky example, because I know you don't. Um, however, if you told me you move to New York during the test done, then I would probably go with that information to update my beliefs about your location of residence. So I think that what you're asking about is whether we're making specific predictions about the different wings of different information sources? And the answer is, we haven't thought about that. And we haven't gotten to the point where we are making some of these predictions. And I think that we would need
a bunch of empirical
work to figure this out. And when I say that, I'm immediately thinking about how the original qaiser at all 2000 small candle debate went. And people said, Oh, are people looking at the small candle, because now they're thinking that the speaker actually knows about that candle? And some of you know what I'm talking about them? And maybe some of some of you don't I apologize for that. Um, so if something is hidden from view, and you seem to, to utter a sentence that should could be interpreted as you knowing about it? Is that my interpretation? And the, the empirical, the, our answer is, we have to test. But thank you. This was cool.
Yeah. Yeah. Thanks
for your thoughts. Yeah, I think it's clear that there's like a bunch of different factors involved with different weights associated with them. So it would be really hard to talk about, in general, because it really depends on the weighting and all of these various things for any specific case. Yeah. Cool. Thank you.
I'm Cindy. Oh, ICU Cindy Fisher going to ask your question that I actually had as well. Okay, now I'm tempted to change my question, because I'm scribbling down ideas for experiments after that last one, but maybe I will share those.
So I wanted to know, so you talked about moving away from the sort of cataloguing and time course questions, I want to go back to the time course question and ask us to sort of
hear you speculate about what new predictions come out of this, if any, I assume there are some about how
about when influential processes about others perspectives, and shared perspectives and unshared perspectives might come up in comprehension
in time, Georgia that online? There is drinking water, so I have to answer this at least first? Um, you,
I assume? Yeah. Yeah.
I'm not very good at indirect cues.
Um, we're not going to
turn on our previous findings and suddenly say, Oh, this is a, this is something that happens late.
We don't think it happens late,
we think it happens right away. But if Previously, we wrote sentences, like, listeners integrate information about common ground in real time processing, then we are probably not going to write this sentence again. So we would probably, I personally think that
we use
both of these perspectives simultaneously. So your own perspectives, your own perspective, and though the assumptions that you make about the partners perspective, simultaneously, in in in real time processing and the relative weighing of those to change by by the situation by different aspects of the situation also by how well you represent the others perspective. And I have over the last couple of years, developed a probabilistic model that that represents that. But what's important is that we can have both perspectives affects real time processing no different than when we had
Common ground Common
ground and assume Common ground representations.
I thought that was good stuff, man. Oh, thank you. Cindy, I'll just
say one of the things I've always been puzzled about by in terms of the time course stuff
came from Well,
in practice noticing, like, we can't write sentences saying windows come around influence processing. And it came from I'm writing when I started trying to write that papers about questions, because I was like, Oh, wait a minute, the questions asked me about privilege grams, so it's not. And it made it hard to write those sentences, because the framework I had been thinking of is was Common ground going
to come online. And
anyway, but but the time of course, thing that I don't have a grasp on, is like, for wh question, the question from English anyway comes at the beginning. And that is a clue that we're going to be asking about privilege round. But then in the freezing of the questions I worked with, you have an NP that is referencing some shared information. So that's what's allowed the pig with a hat. But then to understand the question, you have to shift attention again to something that's privileged. And and I think there's some really interesting questions there. The I can tell you that at least in the eye tracking studies I've done you don't get people looking like straightaway to privilege when they hear a question word, and I don't know why that is. Is it because it's slow? Was it because
there's no content there?
I don't know. I've always kind of wanted to know, but
I got nothing.
Maybe we can take a few more questions. Many assignments is.
Yeah, it is here. And
yes, we can also, like, close this off, basically was the last question for Mandy if she's still here. Yes, I am here.
I code this thanks for a really interesting talk. I enjoyed it a lot. Lots to think about. And
I wanted to make a comment, to comment rather than a question. So
I'm thinking about this from the perspective of presupposition. And I think that the the approach that you're suggesting
potentially resolves a sort of puzzle about presupposition. So, as everybody here knows the standard take on presupposition triggers is that they are supposed to be constrained to be used only when a certain bit of information is in the comment. grand grand there interlocutor is trying to maintain a model of their own their interlocutors knowledge state,
then it makes
sense to me to be sending signals about bits of my background beliefs that on actually part of what I'm saying. And
this is true, not just about presupposition, but about mothership content generally. So suddenly, now we have like a communicative reason for these, these signals to be the
in
in linguistics productions. So I think, Mandy, what you're saying is that, um, if if we are aware that people are that our partner is using, I mean, not aware, I don't mean this in a consciousness kind of way. But if if there is a part of the linguistic system is such that
our partners
are building are using our language to as cues for building the representation of what we do and not know, then you're sending those cues, it makes sense to send those cues.
Besides,
it would be really interesting to look at weather phenomena that have been analyzed in formal semantics and pragmatics as dependent on
like a set of
propositions kind of common ground, can fall out of having just the reverse teleprompter and separate representation, non linguistic representations. Of course, if we're going back to the philosophers, original proposals, they were not generally just concerned with linguistic information per se, they were concerned with more information. Information is information is information without sensitivity to the sources that this information comes in.
That actually
takes us to Dennis's question. But yeah, thank you.
Okay, and for the final question, with about three or four minutes left. Let me make sure I'm getting this right on a Samuel cinema. I hope I'm pronouncing your name close to correctly. Oh, yeah. Yes, the name synonymous, but close enough. Thank you very much. So I'm interested in studying. I just started my graduate program at Villanova. And looking at personality and linguistics. And I'm very interested in how Common ground may interact with a Theory of Mind primer. And using that as a manipulate and having participants run through a Theory of Mind primer where then they go through a Common ground task. Specifically, presser brown Schmidt, you you reference this on your paper as krausen weinheimer is 1966 paradigm which has been since refined but looking at essentially as time goes on, for these number of references, abstract shapes, participants use less than less words to describe these abstract shapes, because they've established that common ground as time goes on. And so I was curious if we had any reason to believe that this you know, Theory of Mind primer would gravitate participants for the more efficient Common ground use where they more quickly use less words essentially, or a faster completion to the tasks. And I don't know if there was any kind of thoughts towards that. But it was really just throwing it out there as a generic
Theory of Mind Timer gaming, like instead of other focus, manipulation to make the person focus on the other person, something like that. Yeah,
so there was an I, I've been struggling with trying to find kind of validated primers for this. But there's something called the imposing memory fast. I think it's Peter kinderman. And it's very, very simple. It's just people read a paragraph. And then they answer a series of questions that grow on increasing complexity, for Theory of Mind require. So starting at, you know, factual based questions like Jenny is a cashier, and then so on and so forth up to you know, Jenny wanted this job to help her family versus Jenny wanted this job to impress someone else. And so it's requiring more and more Theory of Mind reasoning as to why the person did A, B, or C, and we're not really interested in looking at how well someone does on that task. It's just using that as, as a means of getting the participant to engage in a theory of mind mindset, and then proceeding to go on to these common grounds. fast. I guess my intuition is that
the referential shortening that we observe in referential Communication tasks. And our framework depends more on this reverse teleprompter, which is not necessarily a representation of Common ground per se. And part of the reason I come to think that short, referential shortening so you go from a very long description of a tangram to a very short one doesn't necessarily require representations of other people's mental states as such, they could instead rely on a very brief representation. But of course history is that you see, an identical learning rate, if you think of learning is the referential shortening and persons with severe bilateral hippocampal amnesia. So these are people who do not have episodic memory traces who you can talk to them and hide under the desk and pop back up 10 seconds later, and they're like, oh, were you a show referential shortening too. And so and actually shown on unit has some data, working with these patients showing you do see some sensitivity to whom you've spoken to in the past, but it's not based on any kind of episodic memory mechanism. And so
I, I don't know that priming or influencing
a person to think about others would necessarily modulate that referential shortening and I might imagine it could affect other phenomena that are more directly
dependent on
modeling what other people know and so things like questions,
if not probably has a different I was just going to piggyback on the
question and say that, um, I think that one aspect of our theory is, is is that it because we are explicitly looking at the modeling of the other of the representation of the other, it brings us closer to literature on Theory of Mind. That is not does not have a language focus. And I think that that is a, that that is a strength. So there is. So one arm, if you if you think about the language under our model, then it's easy to characterize what you were talking about is a task where you make the representation of the other salient. And then you're putting people in a task where they are going to have to use this representation in, in real time in in the context of language. So I think that that's a that's kind of a shift in how we think about what's different types of interactions, what representations different types of interactions rely on and
it could very last word urgent.
I want to ask a question related to evidential ality unsurprisingly, so I was I was actually thinking about you know, whether it's, it's actually a particularly good evidence for your accounts, or whether it's sort of neutral. Because
we'll know, you know,
you mentioned those
languages that grammatically encode evidential duty, they primarily tend to encode like how information is known, but not necessarily what it's shared with the interlocutor or not and, and I know that there are like, languages with like, super detailed systems that encode like different types of inference, different types of hearsay, and they may also overlap with,
you know, but the casing with a
mouth just shared or not. But these kinds of languages are really, really rare. And we don't really know much about the language use in these cases. So, for instance, I was thinking about the Japanese example. So in Turkish I also wouldn't is the director of that show, if somebody is sitting next to me because I only use the direct evidential for past events, actually. So this is actually like evidential. Also ankles, pants, and so on and so forth. So, in this case, I think it's kind of difficult to tease apart. You know, What that categories for So, um, I guess, um, you know, like, it's going to be more of like more of a comment and a question kind of thing. But I was also thinking about another case in Turkish about the demonstrators, which I think is a much better example of this. Because we have a three way demonstrate the system that knows not only in code, you know, things based on distance. But also depending on whether the listener is attending to what you talk about. So I think those kinds of, you know, ways of conveying shared versus armchair knowledge is actually a much
better support
for this account. Because, you know, I think this theory seems to be accounting for like a broad range of phenomena but
You know, the
thing that is relevant from evidential is actually kind of small and you know, something that we don't see crosslinguistic very often and so on. So Yeah,
I just wanted to
hear your thoughts about that.
So, um,
I will say, um, three things. Um, one is that the reason we, um, got the ideas to start thinking about evidential in this context has to do with examples like the knowledgeable dentist, where once we started thinking about mapping out people's sources of information that kind of like immediately had a little light bulb in my head, go Wait a minute, um, there's there's actually linguistic reflexes in that. So that's one thing.
So that's kind of like one
reason how this is here. Of course, the relevance, particular relevance that we were presenting this in has to do with assymmetries. And those assymmetries according to
the work that I have read and
Korean speakers that I have consulted with Those assymmetries are seem to be there and in fact, not even Kim and I, I saw her somewhere here are in the process of developing a study that will collect empirical evidence to see whether those assymmetries perspective that are expected are in these observed in how people talk. So ask me in six months Um Um Um, I think the The thing that is a puzzle and I will Just say what it is and then give the memory specialist here. A moment to talk about that
