Andrew Piper: Computationally Modeling Collective Narratives.
Andrew Piper, a McGill professor, describes NarraBench, a benchmark for testing large language models on narrative tasks like extracting agents, events, and morals from stories. He's building a pipeline to find collective narratives at scale, but faces challenges with bias in AI outputs and the need for human validation across multiple languages.
Chapters
Hello, this is Kooder Open Lab seminar on November 3rd, 2025. And in our continuing, I think now 102 Cultural Data Analytics All-Star Series, we have Andrew Piper, who is a professor at McGill University. And you're also the chief editor of the Cultural Analytics Journal. And he will speak about collective stories using large language models for large-scale narrative understanding. Andrew, the stage is yours.
Thanks. Thanks for the invite to come talk to you all. I really appreciate it. I'm a big fan of what's going on at the Open Lab. You do all do really interesting work. So this is a nice opportunity to share work, as I will As will become clear, this is very much work in progress. So this is for me a kind of next analytical step in the work that I've been doing, and other folks have been doing in the field on computational narrative understanding. So that's kind of the framework that I'm coming from, and I'll work through this concept of collective stories and the modeling challenges that come with it and hopefully some of the affordances that come with it conceptually and theoretically. So let me, let me just dive right in here. Okay, let's try another button. There we go. Yeah. So if we just ask this very simple question, you know, why do, why do people do things? And why do they make financial decisions? Why do they make political choices? Why do they feel well mentally? Why do they they experience joy, why they believe information, have a sense of social belonging. There are obviously a lot of explanations we could come up for that cause these things to happen. One obvious one is stories, or narratives, more technically stated. So, you know, we— there's— for those of us who are coming from a humanities background, there's sort of, you know, this is totally obvious. I think for other fields, as I'll show in this Next slide. They're just beginning to sort of tap into the explanatory power of narrative. I promise this is the only n-gram analysis you'll see today, or ever, probably, for me. But I thought it was kind of a useful just heuristic to see that, especially in the social sciences and health sciences, that narrative, at least as a concept, is gaining considerable traction over the last decade or 2. And this is, you know, as someone coming from a literary studies background, this is like really exciting. You know, this, this says to me that people are tapping into and realizing the explanatory power of narrative, and that suggests there's a lot of opportunity for, for theoretical discussions, for modeling discussions, and so forth. And so a lot of my work has just been focused on this concept of computational narrative understanding. I'm obviously not the only one working in this field. It's very vibrant. area. You all are working on this area as well, and it comes with a, you know, a range of different frameworks and modeling choices. These are just some of the works that I've participated in, in terms of especially trying to get some theoretical grounding, and I think that is a really important step in the process. You know, for those of us coming from narratology, There's a really long tradition of theoretical frameworks to try and think about storytelling, and I think, as it's gotten traction within computational fields, doing that important work of aligning computational methods with narrative theory has been really important. Our latest effort in this vein is NarraBench. This is sort of an evolving framework for narrative benchmarking, for assessing language model performance and behavior when it comes to narrative understanding and also narrative generation. And so we just, you know, kind of take that earlier theoretical work and apply it to a benchmarking framework. And so the goal of that is really just have that be a benchmark of benchmarks and really encourage people to start to develop. We try and create this large taxonomy that people can contribute to and, and see where the missing areas are. And as we sort of show, the, the biggest deficits right now are not in the dimension of story understanding, but far more in narrative discourse, narrative perspective. And so That hopefully will kind of begin to evolve as a collective project. You know, language models are increasingly a really crucial dimension for narrative understanding, how they behave, what they're good at, the, you know, the so-called jagged frontier. There's just a lot of exploration, a lot of validation. You know, this is a good example of why AI should make slides for us, because I just sort of lazily copy-pasted some of my Google Scholar just to give you an idea of the kind of work we're doing in my lab. Around the concept of narrative understanding and around how language models can be used to push that forward. The basic kind of takeaway on my end is it just allows for a lot more conceptual depth than we have been able to get at in the past with traditional NLP methods. That doesn't mean there aren't challenges, biases, and things we need to look out for and deal with, but I think there's a huge opportunity space there. So if we're going to talk about collective stories, we want to talk about stories in general. What is a story? What is a narrative? I kind of put forward this, you know, sort of interestingly, we have a lot of colloquial definitions of, you know, what are the minimum conditions of a story, but they never really even formalized. So I sort of put this forward in that original EMNLP article as a framework, and it kind of rolls off the tongue if you need to explain it to anybody. So, you know, a story is when someone tells someone somewhere for some reason that someone did something to someone somewhere at some time for some reason. And what that means is, right, we need for there to be a story, We need a teller, we need a medium, we need a recipient who may or may not be present. There's some social situation of the telling, and then there's a motivation or message or value that is driving why the story is being told. And then if we drop below the double lines, we obviously need an agent, at least one action or event, although that's disputable. Some— a lot of narrative theory holds that you need to kind of, you know, at least 2, some kind of like change of I think that arguably can be encapsulated in a single event. Possible object, although you don't need it. Location, timeframe, and then the motivation of, or the cause of the actions involved. In other words, why did this happen? And I don't have the stock examples I like to use, but what's really crucial here is these all need to be implicit, but they don't need to be explicit, right? So the famous short story always attributed Ernest Hemingway as the, you know, the greatest shortest story ever told. For sale, baby shoes, never worn, has very little of this information explicit, but it is all implicit. There is someone telling the story, there's a medium, there's a recipient, etc. And so we can kind of use this to ground our building up towards collective stories. You know, for the narratologists in the room, right, we refer to below the double line as the diegetic dimensions. In other words, everything related to the story world. And then that would obviously mean that the narrative perspective, the narratorial real dimensions would be heterodiegetic, everything that's outside of the narrative frame. And obviously that can cross over if you have a first-person character-bound narrator. The core dimensions here I'm going to pull out are the who, what, why. And so we want to focus on, you know, who does something, what it is that they do, and then the why is not why they did it. Sorry to, you know, add the object in there. The why is not why they did it, but why is this being told? Why is this event being narrated? And that kind of coupling we'll see. So we'll refer to this bottom part, this bottom triple, as an event, which is very standard in NLP literature. And I'm gonna refer to this top portion as the story moral, just to draw on a very colloquial concept from narratology, but that I think has a lot of explanatory power and is very underexplored. And you'll see we've done some research in the lab to begin to flesh this out. But the motivation here is why is this person choosing to tell a story about this person who did something with or to someone? And in other words, it really foregrounds the selectivity of narrative, right? Narrative, you know, like photography, we don't believe narratives are realistic, they're perspectival. And so this gives us this what I'm gonna call the actor-event-value schema. And I do put this reference down to Tim Tangerlini's work. I haven't looked at all of your list of speakers, but, you know, I'm assuming he's either already come or will be coming. You know, his work is really interesting in this building out. He focuses a lot on conspiracy narratives, but has this kind of very strong interest in structural narrative theory, given his background in folklore studies. And so we can, you know, again, when we talk about events, at the center of that is going to be the actor, the agent, and then that tuple there, agent-action-object, is going to constitute the event. And then the story will simply be the sum of these tuples of actors, actions, and objects. wrapped in some kind of moral message. Why, why is this being told? What are you trying to get across? What's the central message? And one of the things I often run into with people is a kind of like, do all stories have morals? And Wayne Booth is great on this. He says all stories teach. And so the— I'm, I'm with Booth on this. I think there is an implicit message or motivation or intent whenever a story is told, and that is recoverable. Can be more or less ambiguous for sure, but that's what this kind of scheme is trying to get at. Okay, so, you know, that's what we're looking at with storytelling. That's what the kind of units that I'm operationalizing here. So what is a collective story? And I'm gonna say collective narratives. We mean the recurring configurations of actors, events, and values that emerge across many individual acts of communication to form shared ways of interpreting the world. Right? So if we want to break that down, it— the idea of a collective story, and I think this is where analytically we move into new terrain, is the collective story sort of resides on top of many individual stories, right? So narratology, narrative analysis, we mostly stay at the document level, right? We mostly say, okay, this document is or is not a narrative or may contain a narrative inside of it, but it is a bounded object. When we're thinking about collective narratives, we're thinking about things that reside sort of on top of that, that unify a lot of micro stories into some larger collective story. The concept of small stories is actually pretty useful here by George Acapulco, where the idea is small stories are these sort of dynamic narrative pieces that coalesce into something bigger. And I think collective story dynamics can be thought of as the collectivization of small stories. So what do we mean? You know, what is a collective story? Well, it's this kind of thing, which is very intuitive. You know, what's the story of Canada? What's the story of COVID What's the story of AI? What's the story of climate, housing, Ukraine, and on and on and on? And crucially, we know that different communities will have different collective narratives that they're telling about these objects. And defining those objects is sort of a first challenge and one that is not entirely clear to me how you zoom in on them. That is to say, how does an object become narrativized? or worthy of narrative attention? How does it become a collective object? So we have some intuitive ones here because these are things we are collectively experiencing, but I think a little more precision there would go a long way. And, you know, just graphically, right, a collective story is the, you know, if a story is the sum of actors and events, a collective story is the sum of stories. In other words, the sum of all these actors and events. Now, What that leads to is a very challenging kind of analytical workflow. And this is, like I said, this is— we're— I'm still at the formalization stage, and I should say we, as I'll get to the end. I have a key collaborator that I've been working with at McGill, Derek Ruths, who's in the School of Computer Science here, and we've been working together many, many years on various projects, and we've kind of zoomed in on this one as sort of potentially worthy of really expanding and staying on top of as a continual effort. So we're very much at the formalization stage. We've done some work, which I'll share with you towards the end, trying to tackle pieces of the puzzle. But it's, again, it's not entirely clear what the units of collective stories are or how you parse them from a big bag of documents, right? So you've got, you know, thousands of documents floating around. They may be multimodal. They may be coming from different platforms. You know, our work in the past has focused, say, just on news. But obviously, with the media environments we inhabit, this idea of a collective story is very multimodal, is very multiplatform. And so that adds a whole nother layer of challenge, which I'll probably bracket for today. Okay, so what do we have for units? We have, right, lots of stories. Excuse me, I always have to mention it. at some point in the talk that I'm a long COVID sufferer, and so my voice often has limited range. And so I wear this headset to help, and it kind of depends on the weather whether my voice is stronger or weaker. Today seems to be a not great day, so we'll see how far I make it, which, you know, I'll appreciate your comments to give me some time to rest my voice. In any case, we have stories. We have many, many, many, many inputs. Okay. And we can either— the filtering of that is already a step that's worthy of consideration. Now, once we have those stories, we can extract people or entities in those stories. We can extract events. I'll list this in more formal narrative NLP terms, but event extraction, agent extraction, and then the moral extraction. In other words, what is the moral of any given story. What are the values that are being associated with these people and these events? And in addition to that, we can think about actor attributes. In other words, the overall takeaway could be some larger kind of meaning or lesson that we're supposed to get from the story. But this also situates the actor within a set of values and obviously valence, right? Is the actor positive? or negative.
You know, the—
it's hard to use an example today without, you know, being politically sensitive, but certainly a character in front of everyone's attention would be, say, the, you know, President of the US. And he's going to be very differently portrayed in different communities' storytelling, potentially around the same events, right? So what you can see happening here is you might have anchors of people, who recur a lot. You might have anchors of events. These are sort of central concerns of happenings or events in the world, and you may tell that event with different actors. You may tell that event with different valence, and you may tell that event with a different takeaway. And so you can sort of see these variables coming together in a kind of graph structure, I think, is one way to think about this, where You know, you've got, let's say, 2 stories here in which the first story is really just about a single person, and some things happen, or they do some things. They are positively or negatively portrayed, and there's a sort of moral takeaway from this overall combination of actor, role, and events, and they have a particular role, right? So a lot of— one of— I'll show you, you know, some of the studies we've done before where you can kind of take the coverage in North America of the conflict in Russia and Ukraine, which depending on where you are will have a very different framing. And we can see how, say, Zelensky assumes a particular role in news coverage with particular valence because he does certain things, and the takeaway might be something very particular. And then in the second story, we have this. But we could also have this be a multi-actor story with different valences, different roles, different events attached to them, from which we have a single, a single moral that we can take away. Let me do a cough-drink pause here for a second, so I will just mute myself as I do this. Give me a sec.
Sure.
Maybe I can mention that, yes, Tim Tangherlini was indeed an Open Lab seminar speaker. One of our greatest moments was I asked him if he knows the Handbook of German Superstition, and he just turned around and pulled it out of the shelf without even leaving his chair. Which was really great.
That, that is, that is perfect. I love that. Yeah, Tim is, Tim is fantastic and, and really doing, doing really neat work. And, and for sure, some of the inspiration behind the modeling going on here is derived from him. And I, you know, I, you could do a closer analysis of, I think, what we're adding. I think one of the, the challenges that I've seen in his work with the conspiracy theory is They get a lot of these nodes. It's a very actor-driven model, and I think it's a little harder with a narrative interpretation dimension of that. In other words, the question is, what is the structure of the conspiracy theory? But that starts with an understanding of the conspiracy theory, and I think the question we have is even more rudimentary and bottom-up. In other words, What are the stories that are emerging? Who are the main actors and events and so forth? So can we see these conspiracy— emergent conspiracy theories taking shape? And can we observe their coalescence? That is to say, how they start to stick and become repetitive. They get told over and over again. Do they morph? In other words, certain actors get jettisoned, certain actors get added, certain events. The moral might also change. as a kind of value system that goes with it. And so I think this gives a nice kind of extra layer on top of the work that they've been thinking about. I'll also look, you know, I think language models really become game-changing for a certain degree with this. So I'll— okay. Yeah, right. So, oh gosh! Hold on. Okay. So, we've got, right, we've got the system. We're starting to extract the pieces with our agent detection, event detection, moral extraction, role attribution, and valence. And next, I mean, there's so many problems. So, this is why, you know, I was like, look, I'm just gonna talk about the formalization. I don't have any results to share with you, because I think it's just a nasty and gnarly problem, but seems really exciting to try and solve, and hopefully will become a collective Whatever. Okay, so now we have a clustering problem, right? So we've got all these, we've got all the extractions of individual stories. So we've moved from, you know, classic unstructured data to structured data. Now they're all over the place. They have all these different forms. And the graph here again is, you know, well, these were definitely AI generated, and I just needed more time to Start to connect the events and the actors, right? So you, you just get this overall network structure of common actors, common events that they're associated with, and common morals, common roles. And out of that mess, right, we've got to start to extract clusters. And that clustering problem seems to me to be one of the more interesting ones. In other words, you know, this— the group on the left is telling this particular version of COVID The story group, the stories on the top are telling another particular version of COVID with actors. I mean, you know, again, the dual positioning so that Fauci, like, you know, one of the challenges you have the same actor, Anthony Fauci, who belongs to different narrative clusters and belongs to different events and values and roles depending on who's telling the story and how it's being told. And the, the biggest problem, right, is what we want to be able to do is say, okay, the— there are several larger stories of COVID that emerge over narrative time. In other words, over the time that we are experiencing COVID, they're going to be shifting collective narratives around COVID. And we might think of this, you know, Intuitively or sort of off the top of our heads, we might think of this, yeah, there might have been one fairly emergent narrative, which is like, you know, crisis, medical emergency, and that is sort of the early experience of it, which then becomes, you know, a very public health-driven story of lockdowns and, you know, depending on where you were, you experienced very different things. And then that you know, gradual fracturing into discordance where you've got certain people very concerned about risk and safety and other people very concerned about freedom, liberty, and so forth. And that interpretive problem, right? So we've got this sort of clustering problem of how do we get these units into coherent subsets? But then we've also got to interpret, like, what does it mean to have all these actors' values And events clustering over here in these roles. And again, this is where I think language models can be so useful, because we can use them at the sort of structured phase, at the extractive phase, but we can also use it as the— as part of the clustering process. In other words, how do we extract narrative structure? Or let me put it differently, narratives From this structured information, right? So we go from narratives, unstructured data, to structured data, to clusters, and we have to get back out something that is itself a narrative, right? That, that I think is one of the more interesting dimensions that language models add as generative systems. You know, if we think about Tangerlini's work on the conspiracy theory, you'll see, you know, Clinton and, and the pizza parlor place and, and so forth. And they're all linked. So we can see who, you know, who's part of the conspiracy from some subset of articles. But we don't really have— we have to know what that story is in advance, right? But what LLMs allow us to do is say, okay, let's, let's now take that set of information and the setup for those documents, and let's extract a large-scale narrative and do a kind of narrative summarization of narratives. And so there's obviously a huge challenge of validation here. But I think what it allows you to have is something that is far more, far more actionable with stakeholders. So that's one of the other, I think, big claims that we're, we're trying to make here is that you get this kind of narratives in, narrative out aspect to this analysis so that you can do updates in eventually real time. You know, the plan is this is supposed to be kind of real-time analytics so that you can give Public health agencies. Look, these are the dominant storytelling structures out there. If you want to reach this community, you have to tell your story in this particular way. If you want to reach this community, and so forth, or the story is starting to go in this direction. If you don't want that to happen, you need to start, you know, creating a lot of storytelling that tries to move that collective narrative or moves people from this group to the other group. And so You know, I think that for us is where one of the affordances isn't just, you know, curiosity or historical analysis, but hopefully has some sort of like social function in allowing institutional actors to engage with these collective narratives and move them, react to them, engage with them, speak to them. You know, I think I just wanted to— when we, you know, when we're talking about morals and kind of narrative values to emphasize for those of you who haven't seen this work. You know, this is something language models are pretty good at. And they're, you know, they tend to be strongest when we're talking about the distillation of large amounts of information, sort of that synthesis. The blandness problem is actually in your favor. I'll talk a little bit about, I mean, what we've done in the lab. I'll show you some papers, and generally what we find when we validate language models' sort of moral understanding, its ability to extract story morals, or what's, you know, what's the moral of the story, is that they, they are within the space of human variance quite centrally and generally preferred. In other words, if you had a bunch of human annotators write what the moral of The Great Gatsby is, or lesser-known things, and then you ask the language model to do it, and then you give these to random voters on Prolific or something like that. Generally, people are going to choose the language model's moral. You know, it's a hard task, right, to synthesize large amounts of information into a coherent statement, and that, that slight randomness that individuals produce actually is controlled for a bit with language models. Now, The limitation of that that we're finding, especially when you go to multilingual contexts, is a multicultural context, is a lack of diversity, right? So the ability to model a kind of like cultural consensus means that you, you have a harder time modeling cultural diversity. So that's a secondary problem that we still want to think about. But in terms of the overall quality or reliability, it's, it's quite high. So this is just an example of The Great Gatsby, classic work of American fiction, you know, and everyone's Like, yeah, but it— because it knows it. And of course, yes, this is true, but we can also test this on unseen stories and get very reliable outputs as well. And so this is the piece where we did that validation. We used both automated and human preference metrics to just sort of ground the fact that language models are quite good at deriving these stories that we're calling story morals. And, you know, we can, we can give these like sort of perspectival hypotheticals to language models as well that again track obviously with our expectations. This is just to simulate like, look, we're— this is the process we're gonna go through, right? Is, is we can give either the structured information or the unstructured documents and try and extract these things. And so what we're gonna start to do is compare in terms of interpretability, in terms of validity, in terms of fit. Do we need that structured pipeline, or is this really just document summarization with a particular goal in mind? And so these are all kind of like, you know, exploratory questions that we're gonna be thinking about. We did do one particular project around a given theme, which is around climate change, comparing North American media to Chinese media, both state-run media and offshore media, to get a sense of the differences there. And this is just sort of an example of like, you throw a random news article and we can ask for very classic narratological components. Who's the protagonist? We could ask for a very simple tripartite role labeling. We've also done open-ended labels that are quite good as well. And we could look at central topics, valence, and then this sort of moral extraction. Now, typically, right? So this is sort of the pivot, I think, is typically what we've done when we've done this large-scale analysis is we'll, we'll just sort of aggregate all of the, the individual pieces as observations and say, you know, what are the common recurring most distinctive components, right? So if we reduce these moral sentences down to keywords and compare the distribution of those keywords across Chinese language media, both state and offshore, and North American news, you see certain kind of different moral concerns. The climate change itself is a much more central concept in North American media, but also this idea of adaptation versus cooperation, development, and this sort of more internationalized understanding of climate is much more common in Chinese media. So that's the, you know, if we want to understand the story of climate change in these large-scale ways, this has been the very typical way we've gone about it. This is an example of how we can use language models to scale up our analytical precision. But I think the goal is really to say, okay, Let's rather than aggregate individual units and leave them as units, start to aggregate them as collectives. And so that workflow looks something like this: agent detection, event detection, character role labeling, story moral understanding, narrative clustering, and then from those clusters, some kind of narrative summarization where again we can think of that as another layer of story moral. What is the moral of all of these stories, or what is the story of of all of these stories in this cluster. So that summarization aspect takes us into new terrain, I think. But again, language models are going to be very useful. Okay, so I'm going to wrap it up there and leave plenty of time for questions. I want to give a big shout out to my local collaborator, Derek Ruth, who we've been doing a lot of work with, the students David Hobson and Haichi Zhou, who are master's students, one in computer science, one in our master's in digital humanities program here. where we get a lot— we get a— not a lot, it's a very small program, but a really wonderful program. We get a crop of fantastic students doing really interesting work in the cultural analytics space. And so if you have aspiring undergraduates who are really interested in a very personalized pathway through cultural analytics, digital humanities, computational humanities, do encourage them to apply, because we're always on the lookout for excellent students. And also thank you to Maria Antoniak, who's a junior assistant professor at University of Colorado Boulder, who I've been collaborating a lot with on narrative detection, and is also a collaborator on this project, which is, you know, making its way through the works of grant applications and things like that. Okay, I will hand things over.
Super cool.
Thank you very much. This was great. So, very rarely we have talks where people actually present sort of like a proposal, and I think that's a really, really great mode in essence, because that is sort of like making the discussion even more fruitful maybe. People were clapping. We have this thing which we learned only a couple of sessions ago is that Zoom actually sort of like drowns that out so you can't see it and hear it. You can hear it, that's the thing. So thank you very much for this fantastic talk. Those in the audience who are here for the first time, so basically the idea of this is we have Now had the talk and now sort of we switch to discussion and ideally raise your hand and then we take you one after the other. Today's session, we will have about 55-minute discussion because Andrew has to leave actually at half past over the hour to get in time to the next classroom to teach. So do I see any hands up?
Yes.
So Mark Metz, and I just sort of like focus the camera to whoever is speaking. Yes, Mark, go ahead.
Yeah, thank you so much for the talk.
It was really, really exciting.
I wanted to ask, so the talk was inspiring, and with what I've seen with how LLM can extract different parts of narrative, and then just to think that all of the different the typologies that you can try to extract from there and which one you can infuse from them. Like, it sounds really inviting, but then again, it seems that I myself have run against this sort of wall that how do I retrieve language models trustworthily enough? And what is— I wanted to ask you about that part. What is your take of this current state and where are we going and maybe Maybe even some tips and tricks that how do you keep it simple enough that to get these narrative elements from moral events, etc., it would be as high quality as possible.
Gotta get my audio back on. Yeah, just to— there was a word I lost in there in the echo. Was it trustworthiness? Is that the term you were using, or?
Yes.
Yeah, yeah, the problem of evaluation, etc.
Yeah, yeah, yeah. Yeah. So yes, I mean, this is, this is our problem, right? Like, this is our, our collective worldwide problem. And it's a really good one. And I think the, the methods we use, I mean, I think you just have to incorporate human evaluation into your workflows. And so that's, That's what we've been doing, and, and, and I, I think there is a place for automated metrics with human evaluation. So, and I'll talk a little bit about the successes and biases that we're seeing. So, because we have a new paper exploring this as well in a multilingual context. So the, the success is that, that synthetic ability of language models and We, we do not so far with the, with something like story moral extraction. We're not seeing hallucinations. I was, I was working on another side project on narrative summarization in order to do story moral extraction. In other words, one of the things we explore are inputs for, for, say, the story moral extraction. In other words, do you need the whole book? Or can a story— can a summary suffice? In the case of Wikipedia, we have lots of high-value human summaries, but that we don't have them for everything we want. And so the question then is, can language models be good story summary generators such that you can extract story morals? This can also be useful if you're working inside of proprietary, you know, capsules and things like that to work with copyrighted material. So there's already, you know, there are these like input questions which we've been exploring, and I'll, practically speaking, we generally find that some kind of summarization step improves the quality of story morals. In other words, you've got this big baggy thing called a narrative, which can be anywhere, you know, very short news article, which maybe then you don't need the summarization step. But once you get to a certain length, asking a model to think through and reduce the dimensionality down to sort of core aspects can really facilitate the moral extraction step. So that, that's also something we often think about is like, how many layers are there to this process to getting the best possible answer? But ultimately, what we're, what we wrestle with is the thing called the best possible answer. And I think one of the things we did in, or we do in Nara Bench, which You know, a lot of people are talking about this problem of cultural alignment, perspectival alignment in general with language models. And the way we tend to think about this is, what kind of task are we dealing with here? Is it a discrete extraction task? Say, who is the principal actor? Or give me all of the actors in this story where we expect there to be a right answer, and we expect there to be minimal disagreement about that. Up to much more interpretive tasks like, what is the moral of the story, where we expect there to be potentially divergent answers. Our experience has been that that space of divergence is much smaller than I think most people might initially assume, or that literary criticism would lead us to believe. I think that, you know, the history of literary criticism is like all about the empowered interpreter, the variance of language, the indeterminacy of language and narrative. And I think that's a— I don't think that's the best model. I don't think that's the best theoretical framework to start with, given our experience of when we ask people to do these tasks, we do see convergence around themes and concerns that recur, but we also do see variation. So the challenge with something like this is you have discrete validity at the level of like extract the actor, And then you have a distributional validity at the level of, say, character role, potentially valence, although I think that's more deterministic. And then certainly with, you know, what's the moral of the story? And I, you know, I don't think we've nailed how to solve that problem, but we always want to make sure we're including that problem. And so, you know, very practical steps are try and use multiple models. Try and use random variation in your prompting structure. You know, a lot of good work has shown that LLMs are very sensitive to word wording, even at the level of like not semantically different, right? So if you just permute word order, grammar, and things like that. And so we do that too. So we sort of try and inject as much randomness into the process so that we're de facto capturing a broader distribution of potential outputs. And so the question is, you know, how many do we need per story? How good are language models at creating multiple perspectives? And, you know, we have time in this Q&A, so I'm gonna just go on a tiny bit more to answer your question about now the limitation. So what we've seen is if we ask humans to do this, and as I mentioned in the talk, the language model outputs will be, you know, if we think of this as just like a semantic space, the language model outputs will generally be within that space. In a particular culture, right? So most of what we're doing here is North American English. And yeah, both in terms of the content and in terms of the validation. Now, with the Chinese study, Chinese-induced media study, we found the language models were very good in that language as well. And so when you're dealing with these kind of very large supported global languages, it appears to be the case that we did not get major cultural variance there, that the validation worked as well with the Chinese annotators' answers. The tiny confounder is that we use native Chinese speakers, but they're in school in North America. So obviously there's some contamination there. But in general, that, that is pretty good. The limitations we find is, let's say you have stories from many different languages. And you expect that there should be some cultural variation. You should not get the same moral, right? If you, if you give it control for the story, give the same story to a Japanese reader, give the same story to a Canadian reader, they— there should be some cultural difference there, right? There should be some predictability about the, the values that a Japanese reader may extract versus those of a Canadian reader, you know, and, and so forth. Again, I think the cultural difference theory suggests that that's way stronger than what we've observed. It's very slight. So we've done a human preference study where we, we permute combinations of story morals, and then we'll ask people from particular cultural backgrounds to rate their preferred moral, and we'll use morals produced by people of the same culture and morals produced by people not in the same culture. And you get a like tiny statistical effect. That annotators seem to prefer the morals produced by people from their culture, but it's very, very weak. Like, it's very close to random. And so this suggests that there are implicit latent signals around culture that influence story moral extraction. They're not super strong, but one of the problems when we use language models is across languages, the, the variation there does not mirror the human variance. So, you know, you should expect to see culture to culture, you know, more fluctuation. And there seems to be fluctuation that is culturally specific. It's not just random people, but that the language model basically has a meta archetype of the moral and it's just translating into every language. And so it's basically giving the same answer. And that's, that's behavior that's limited. So when we're getting into multilingual contexts, I think we have much greater hurdles than when we stay in this sort of predominant single language context, which is where we're, where we're starting. Anyway, long answer, lots of, lots of pieces there if you want to engage with it, but I just wanted to give you some idea of like the different strategies we're using and challenges we faced.
Yeah, thank you. That's, that's, that's, yeah, what I was looking for.
Thank you very much. Um, Jana Sachi has the next question.
Uh, thank you, Andrew. Great, uh, great talk. Probably should have read more of your papers before asking the questions, so they might might sound like obvious if one reads everything you read, but you wrote, sorry. But I was about to ask 2 things that you touched in your previous answer as well. One is maybe the overrepresentation of the loudest voices in literature and in media. And should we as researchers be afraid that that, you know, someone might hijack that for any particular reason. It used to be in the media, like newspapers and stuff, where, you know, the headlines would converge in opinion. Social media would produce specific content for, for users to observe. And now LLMs would produce answers that are kind of tilted towards a desired narrative in a way. And also considering how much we can produce now using LLMs, so we can make quite a lot of slop in a way. How, how do we as researchers protect ourselves, you know, and make valid claims about the culture and what is—
Yeah, yeah, no, it's a, it's a, it's a good question. Um, I, you know, I think there, there are 2 things that, um, are going on here. One is, for better or for worse, when we're talking about collective narratives, The, the loudest voice problem is not a problem. In other words, it's the thing we want to capture, right? So, you know, our, our challenge as citizens is the loudest voice problem that we have a media environment that rewards conflict, that rewards outrage, et cetera, et cetera. But as analysts, oh my God, as analysts, As analytics, as analysts, we want to observe that. And so we essentially are tracking that space. And, and this is where data collection comes in. And, and for reasons that I can't explain, I left that out of the talk. But there, for the, for the initial pilot of this, we're using— there's a Canadian Media Observatory located at McGill who should also be on my, my thank you list. My ability to leave crucial information out of slides is legendary. So Taylor Owens, thank you. If you ever watch this video, I really appreciate the work you do. He is a wonderful prophet at McGill who works on Canadian media and media studies, and he runs this Canadian Observatory. And so it's constantly collecting, and Derek Roots is a participant in that, and they're constantly collecting this sort of Canadian media space, right? And so We expect it to be dominated by the loudest voice problem, and in some sense that's what we want to capture. Now, if you're doing your job as an observatory, you are also making sure you have a very diverse array of media that you're capturing. And so I think where that clustering task comes in and its strength really shows up is, yes, there are going to be predominant narratives, but there also are going to be minor narratives. And I think That's something we were really thinking about in our, in our framing of this project, which is, how do you maintain that minority voice within, within this larger loud, cacophonous narrative space? How do you discover it and make sure you're also shining a light on that, right? So some of that comes from data collection. You, you really need to make sure you're not just getting the algorithmically driven stuff, because that's— you get your loudest voice problem. So you have to sample well from the media space. But also, I think this methodology allows you to pull that apart and say, yeah, there, there's something different going on over here. It's very different messaging, and it's not consuming much of the ecosphere, but we want to hang on to that community's concerns, right? So again, if you're a public health institution and you're thinking about reaching different communities, you can think about our system. You know, you can target their platforms that are more vibrant for a given community and think about, okay, what's the narrative that's emerging there? So hopefully we can sort of do both things, and it's really a matter of, uh, of perspective.
Right, thanks. That answers my question.
Thank you very much.
Um, Leonardo wanted to ask a question.
Yeah, already? Nice. It's the next one.
Uh, yeah, sorry.
Uh, I wanted to ask, uh, on, on the answer you gave previously that you mentioned that culture is one of the kind of determining factors Uh, when people interpret the moral of a story. I was wondering if you found, uh, other influences such as, for example, class, since we know that for the class can influence taste heavily. I was wondering if it does the same thing with interpretation of a work. And in general, what do you find that these, uh, collective narratives cluster around when it comes to meaning, if they're the same marker, markers that are usually associated with identity?
Yeah, good question. We have not, we have not done that yet. In other words, the only demographic information we've, we've conditioned on is, is native speakerness. And so limiting, doing a potentially like single culture around a more vertical understanding of identity, education, and so forth. forth we haven't done is, you know, a really good suggestion for our next step. I will say, this is not in our defense, one of the biggest challenges is that these human studies are ridiculously expensive. And so for a lab like mine where we're, you know, operating very generously from Canadian federal grants, but they're in the social sciences and humanities, which are not large. And so they're very like individual researcher targeted as opposed to large-scale, you know, horizon-level grants in the EU. It makes this kind of large-scale human preference questionnaire very challenging. And so hopefully we've, you know, we've been applying for different kind of more collective grants to think about supporting that work using something like Prolific. There are enormous challenges, obviously, Now with the reliability of human-generated crowdsourced data because of AI. And so the reliability of platforms like Prolific seem to be, you know, I wouldn't say collapsing, but it just becomes a huge problem. We even with our, we had all these checks and balances and in just making sure, you know, was someone a native Japanese speaker when they took our survey? And we have to— one of the classic checks we have to put in is if their answer sounds too much like the AI's— an AI answer. So we have to get a bag of AI answers, and then we have to discredit any that are too close, which you can see becomes a bit of a confounder if you're trying to figure out the human-machine similarity. Because if they're using AI, they're gonna sound like the AI. If you discredit any answers that sound like AI, then they're not gonna sound like the AI. And so a lot of it just becomes manual gut checking. You know, we use similarity thresholds, but then you, you, you really have, you know, my take is you have to be extremely cautious with assessing human preference data because, you know, either you need a scale which is just massively expensive, or you've just got so much noise and unreliability. I mean, we talk about the trustworthiness of LLMs. We have a huge problem with the trustworthiness of crowdsourcing. And the only thing I'll add to that is we've been doing a lot of citizen science work, which I have been super excited about because it solves some of these problems, not all of them. In other words, if you're asking people— so we use Zooniverse and we get really nice participation of, you know, we've had, I think, over our 3 projects, we've had 17,000 volunteers. And, you know, well over half a million annotations and data points and on literary annotation. And so it's been this really wonderful awakening of like, you know, citizen science is a very robust area. Zooniverse turns out to be a wonderful platform with a lot of enthusiastic users. And our hypothesis was that, hey, there are a lot of people out there who love to read, love stories, and maybe we can tap into that subpopulation of Zooniverse users who like to annotate animals and galaxies, but might also like to annotate for books. The limitations are, we know very little about them demographically, so we can't do exactly the kind of questions you're asking. And B, there's— the tasks tend to be— have to be much more limited, because it's geared more towards this short task thing. So if you want to read something long, it's not as good a platform for that. So So, so it, it solves certain things in terms of the validity and reliability. Anytime we measure their agreement with our local moderators and annotators here, we see very, very good alignment. In other words, the quality of the data, and you can sort of see this from an incentive structure. You're not getting paid, so you're not trying to game it. Like, why would you come in and use AI to answer questions voluntarily? Seems like a really weird state. Obviously, we can still check for it, but that, that's one of the things that we see. That volunteer participation motivation gives you much higher validity and reliability. So to answer your short answer to your question is, yeah, we really need to test for demography, and again see what those differences are and what kind of material differences, what kind of texts produce the greatest variance, you know, news stories, etc., etc. So yeah, I think there's a lot more to do there to understand the human baselines. And then compare those to the LLM, knowing that LLMs are very monolithic. Their reasoning space is very narrow. And so knowing those human baselines of variation will help us understand how much, like, we need to push LLMs' creativity and diversity. But the central problem remains of that LLM homogeneity.
Excellent. As you can see, I got a scary question. But first, Chico Camargo, actually, has a question from the audience.
Hello. Just continuing on the citizen science thing, I love that you mentioned that because I'm a big fan of approaches like that. I've been trying some of my own, and there's— well, then you run into the trouble that you've— your participants, your annotators are people who like annotating stuff, right? Which maybe in this case would be okay. Could you talk a little bit more about like where you see— so you just started to say, you know, like it works for small tasks. Could you see a citizen science approach working for anything beyond just annotation? Or, yeah, or could you see any issues of any biases that might come because of this type of person you're selecting? Because for comparison, I did citizen science, for instance, up in a study where the theme was quite political. So in that case, the way we found our volunteers actually definitely biased the results, but we were kind of aware of that. I, in a case like this, I'm really curious about what would even count as a bias.
Um, yeah, can you just talk about that? Yeah, that's a, that's an interesting question. So yeah, and I hadn't thought about that of Yeah, people who participate in, in Zooniverse are people who like to annotate, you know, which is a particular kind of person, right? Not everybody likes to do that kind of work. So, I mean, the only thing we, we can do at the end of the day is we, we measure inter-annotator agreement and we either compare that to our local. So we always have a team of— one of the really key dimensions of citizen science is moderation. So it's also very different from surveys where you're actually interacting with your population, because again, they're volunteering their time. You want to respect them, you want to engage with them, you want to answer their questions and clear up ambiguities. So it's— it is ultimately a very different set of outputs. I think it's probably closer to, you know, it's just a better version of Mechanical Turk would be one way of thinking about this, where you're going after— I think there's approximate set of answers here that we, that we can get, and we just need many eyes on the problem who we can sort of trust. I think if you're really trying to do more survey-like stuff and understand differences, then, and I think, I think that's a severe limitation. We did an emotion annotation task on narrative emotions, and I think that was for us the absolute boundary of like where I was comfortable with subjectivity. In other words, That's, that's like, what is this character? What emotion is this character feeling? Is somewhere between objective and very subjective. And so there we really amped up the overlap, right? So one of the things you can— a parameter you can manipulate is how many people have to annotate each task. And there we, we wanted to have at least 10 people doing each one to get a broader cross-section and see what kind of consensus emerges. And so we do inter-annotate agreement, compare that to our moderators. So we always have our moderators there. They're the ones who test out the projects before we put them up on Zooniverse. I mean, Zooniverse is amazing. It's such a— you can really— it's like a very easy-to-use platform. It allows you to iterate. They have beta testing in place so you can really test out your projects before they go live. You get really detailed feedback. We always test locally, obviously, with our moderators who are undergrads who've been working with the lab for a long time. So we get this kind of like shared understanding of our goals and the boundary cases. And then we kind of use them as the gold standard, right? So that's, again, We, we like to see if these unknown people who have this bias towards liking to do tasks, do they look basically similar to the most trained people I could come up with in, you know, in my proximity? And so far, we've always seen that. We've, we've, we've never really found in our comparisons strong divergences there. And so the best that I can say from our validation, which I think you could do you know, much more rigorous deep dives into this. But the validation we've done for our projects has shown that the kind of data we're getting approximates what you would get if our students, you know, could do everything, which obviously they can't. So yeah, it's— I really support, I really encourage researchers to try it out. I think Zooniverse is an amazing community and what they're able to do, the volunteer community that they have is amazing. I think experimenting with you know, lots of coverage on smaller number of things just to get a really broad swath of answers, even if you don't know a ton about the people who are answering, is interesting. We're finding you the scale you can get is almost like you don't need language models. Like we sort of started out like we're just going to annotate a bunch of stuff and we're going to train AI and we'll— this will be a more transparent process of like people-driven AI. And like we're getting such huge scale that we can just do the study on the human annotations, right? The AI becomes a, you know, we can use them to benchmark AI, but we don't even need to worry about LLM trustworthiness because we've got this very reliable, active community of human annotators. And that, that was kind of like a, whoa, that's an exciting new opportunity. So I, I proselytize a lot about the beauty of citizen science and Zooniverse and really want people in our field to, to, to use it more if they, if they're interested.
I think there's a, there's something in what you mentioned there as well, that all of these things that you described are, or a lot of them are still using the notion of a gold standard to using interannotator agreement. You're trying to get consensus and blah, blah, blah. But actually, the fact that you're using a lot of people with different lived experiences can also be explored in the diametrically opposite direction of trying— not trying to see if they agree, not trying to get them to agree, and not measuring that. I'm finding that lately, I don't know, I've been thinking about a lot of this lately, of how we have a lot of good measures for this paradigm of trying to get the right answer or or the right answer according to, you know, stuff like if, I don't know, Andrew likes pizza and like what's his favorite pizza? Like there is a right answer if you have a favorite pizza, even if it's subjective. But there will be cases where people will have different opinions, as you've been pointing out, and those both are valid and an inter-annotation score will make it look like it's a bad thing, but not quite. And I feel like there's a lot of work in this funny nexus of computational humanity where I feel like we need more measures of— I'm not even quite sure why, but, you know, of like the diversity of perspectives and blah, blah, blah, because often we're interested in that. And so much of the machine learning work is, is geared towards gold standards and ground truths and stuff that agrees. Anyway, I'm not going anywhere with this. Thanks for the answers.
Yeah, no, I mean, just to— yeah, you, you make a really key point. And there is, you know, it is becoming a much bigger topic in NLP, is— and I think that, you know, NarrowBench for us is really the place where we're putting a lot energy to try and figure this out. In other words, we, you know, we survey, I don't know, 75-ish narrative-adjacent benchmarks that exist already in the field. And the 2 big biases we see are one, most of them are around story content. So just extracting agents and things like that. And most are discrete answer-oriented. And that's true of benchmarking in general. Q&A, there's like a right answer. And as we know, right, when we study narratives, that's, interpretive range matters. I, I will go on record on YouTube as saying I think it's always narrowed narrower than our theories predict. But nonetheless, it does matter, and coming up with metrics and approaches that incorporate distributional answers as opposed to single answers, discrete answers, is going to be, is going to be really key. We did this with the Reading Emotions Project, where you can get a lot of answers through discrete emotion words, which is fascinating. And then the way to sort of build that out as a metric is to map them onto, say, valence, arousal, and dominance dimensions, where then what you get is a distribution of answers for any given sentence or a distribution of answers for a collection of sentences. And so what you're saying is like, there's not a single emotion here, But, but people feel like this is what the range of emotions looks like. Again, it's not anything goes, but it's also not one answer. And I think that can be useful for, for how to incorporate that into analysis of phenomena is also a next-level challenge.
Cool.
Before I ask my scary question, I got a follow-up to this one, which is, you know, in some sense, the large language model already runs on data. which is produced by many, many people. And if we consider sort of like the pipeline of like putting a prompt in and getting something out, a little bit like a telescope, and, you know, something happens in the tube, you could imagine the situation where Galileo points his telescope on the balcony in Rome to the moon. And then having basically the crowdsourced sort of like balance of that would be like asking random people on the street, letting them look through the telescope and describe what they see, which comes out of their head. It does not come from the moon, right? And so that doesn't make it more true. While at the same time, if we point the telescope at different places, you get the same kind of diversity. So the question is, do you really need real people? One of the most sobering experiences I had myself with this, I was asking Amazon Mechanical Turk users, what are your 3 favorite classical music composers? And I thought this was like sort of, you know, a 2-hour afternoon kind of thing. On that day, I learned a lot about Indian classical music, and I still value that a lot because I learned a lot of things that day which I would have never encountered because you simply— there's no record store and no section in Spotify where you can actually sort of like find that kind of stuff where I come from. So, but one of the key issues is that, you know, it's important to know who is these people who actually answer these questions, what kind of questions do we ask, and we get this sort of like art form of like constructing these things, like just like sort of constructing surveys is an art form, doing sort of like crowdsourcing and citizen science. You know, if we— if you look at Luis von Ahn, the success of Duolingo, the success of reCAPTCHA, like all that kind of of stuff is sort of like born out of this. But the question is, is that something which we can use and say, okay, so I want to know the bias of my large language model, and therefore I use citizen science, which is like, aren't we sort of wrestling 2 octopuses at the same time? And isn't that dangerous? Basically, that's my question.
Yeah, it's a really interesting point. I don't think it's dangerous. I think it's You know, the best we have of 2 problems. And in other words, you know, the taking humans out of the loop just seems too, you know, speaking of scary, like, just seems too scary to me. In other words, language models can do weird things. And so making sure that we have some checks and observational, like, feedback loops, I think it certainly at this stage of the game is still very important.
Yeah.
And I completely agree with you that, and this is what we generally see with the kind of tasks we've been engaging with, is that, as you said, language models are themselves representations of distributions of human responses. And that means when you get something out, that synthetic power of a language model is quite powerful because it's doing that for us. And so When we're asking a subset of humans, yeah, we might just be finding a portion of that distribution that the language model has actually access to a broader distribution, which I think explains why we generally find humans prefer language model answers on a lot of these types of tasks, especially with story morals in particular, because it has— it finds that sweet spot. It's that telescope which has just got a, you know, more magnification at its disposal. Um, does that mean we're getting bias if we use either citizen science data to train a language model or interpret a language model? I mean, 2 other things I would want to think about, which is the, the problem of language models in their current state is obviously the reinforcement learning layer and then the system prompting layer. And what they're doing to that capturing of the broad distribution. In other words, they're being systematically biased in their sort of packaging towards public consumption. And so we, when we're getting these like human referendums on language models, we unfortunately, well, not entirely true. In other words, we have to work to go back to Seeing what aspects of the model, pre-training data, instruction tuning, reinforcement learning, et cetera, et cetera, are pushing that telescope in one direction or another. And so, in other words, I don't think it's dangerous to use one octopus to study another, but they are going to tell us about each other. And so I think, you know, I think both aspects are important. Neither is ever perfect, right? It's just like really hard in today's age to get like a sample of human beliefs. And so, as we move out task complexity, obviously we get more ambiguity. But I, you know, on some level I'm just like, well, we're researchers. This is what we do, right? We incorporate ambiguity and uncertainty into our estimates. We draw conclusions, and we try and understand just how much variability there is in that space, and how that might impact the kind of conclusions we're drawing about a given task. So I don't know if that's satisfying or not, but That's how I would, I would think about this issue.
So, um, here's my scary question.
Um, for the record, before you start, that pumpkin is not scary. Just, it's a very cute— it's a very cute pumpkin.
Yeah, exactly. Can't even do this.
But go ahead, sorry.
I, I also— I didn't do it. It's like, uh, the crowd who had a Halloween party. Um, I did leave it in the room. So I'm also ironic in terms of scary question, but also only half ironic, actually. So the key thing is that, are we behind the curve in trying to understand conspiracy theories when actually there is a conspiracy going on? Like, to give you an example, that's very low level, level, like the Ronald Reagan video, which, you know, the Canadian advertising where he said, like, actually he doesn't like tariffs. We're awash in these kind of like counter-AI videos, which sort of like try to sort of like obfuscate that he said that actually like that. And even if you're very trained in sort of like getting back to the sources as a historian, you will have a hard time. actually sort of like validating that Reagan really said that. So, but what's rolling out currently politically, and I think it's very hard to not be political in that kind of sense, is sort of something that has a plan. And you could say it's a conspiracy against open society. So are we a little bit behind when we sort of like trying to understand conspiracy theories and like, you know, sort of like neo-kayfabe, the sort of like politics version of the World Wrestling Foundation, while at the same time we should really look at like what's the kind of rug that's pulled out and how it is pulled out under our feet as science is defunded, as sort of like rights of women, LGBTQ, civil rights, immigrants, and so on are taken away. And, you know, that's not a conspiracy theory anymore. It's actually happening. So, so, so, so what is our role in this kind of thing? Um, are we doing more of the kind of stuff we did the last 5, 10 years because we all want to get tenure, more funding, whatever? Or like, what is our role in society? So that's the scary question, basically.
Yeah, that is a lot scarier than the pumpkin, I'll tell you that. No. Yeah, I mean, it's dark. Let me think. So, let me walk, let me just give you the sequence of my brain here. So, the first step is if we say, okay, imagine this system works, right? Like, let's get past all the mechanics we've been talking about so far and just say, yes, we have this Canadian Observatory, a news observatory. We have this real-time workflow using language models to extract clusters of collective narratives. We can locate them either thematically or by community. And now we have a real-time landscape of the stories people are telling in society. And I think your point is that that's just like, that's the smokescreen, right? That's the like, look over here problem while we like take all your funding away and, you know, blah, blah, blah. Yeah, I mean, on one level, I would— I'm just sort of falling into the trap of like, feels like a separate project to me, right? Like, like, you know, you know, I'm gonna like, I'm just gonna spend a— unfortunately, I can't think about this while I'm teaching, and if I do, I'll be a very bad teacher today, but I will be spending a lot of time thinking about this. What is the derivation that we can get there? What is the critical information that we can bring to light to address the problem you're talking about? And the question I'm asking myself is like, is it also narratively structured? In other words, so-and-so. So, you know, we're extracting statements like forms of expression, right? All the documents that get spit out and may in the future be AI slides. So what is the point? I'm rehearsing my own demise here. What is the point of extracting collective narratives if these are just the collective narratives of AI slop and they're not really anchored in reality? Who is, who are the agents that are doing the actions that are having these results, right? That, that would be, I think, another way of thinking about modeling it. And we're capturing the simulation and not the underlying thing. The optimistic version is that power still needs stories. And I don't know how true that is, but that's the argument I'm going to go with, which is, in other words, the collective action or belief or politics around COVID, I think, is not simply institutional actors behaving at will, but there is an interface with the story structure around that experience. And so when a particular agent is demonized or when a particular technology is demonized like vaccines, you know, on the one hand, the director of a health system in any given country can just pull it and say, you can't get this vaccine. I don't believe That project gets a lot easier if there's a widespread belief that this agent called a vaccine, this agent called a vaccine, does these nefarious things. And the moral of stories of vaccines is that they cause autism and all sorts of terrible consequences. So, you know, my hopeful answer is that these acts of power, the conspiracies that you're talking about, which of course absolutely exist, right? This is, this is where my concern with Tangerlini focusing, you know, Tim focusing On conspiracy theory as a starting point, as this like fantastic realm of narrative. I'm like, well, I don't know, right? I don't know where that boundary starts and stops of when it's true, when it's not. I just want to map the narratives. But I think, you know, my hopeful argument is that power and decision-making and institutional actors, they need storytelling. They need that to guide collective beliefs in order to do what they do. How true is that? And in the most extreme cases, does Putin need anything to do what he does? You know, right? I think there might be a spectrum there where this theory breaks down. But I think in democratic, quasi-democratic societies, I think there is a linkage between institutional action and story structure. But that also might be an empirical question, which we can help hopefully answer. I don't know. I'm curious your thoughts on that answer, um, and how you see this playing into narrative analysis. In other words, are we just looking at the wrong thing completely? What would that other thing be?
If you imagine that, uh, we have the actual story, like we know where the money comes from, we know who is actually guilty, like imagine the people who are now not visible. but they're guilty and we know it and we have that story. We can't tell it to anybody because if we tell it, people will just say, oh, that's just another opinion, it's kayfabe, right? It's just like, you know, in reality Hulk Hogan is a nice guy, right? So you see what I mean? So, so the key question is like, even if you know, like imagine sort of something like the Nuremberg processes, like how did they establish truth truth of what really happened? That is a really good question, I think, because in the end, that is what it boils down to. Who will ever be put to jail or whatever for what happened in Mariupol? What happened in, I don't know, on the music festival in Israel and in Gaza? How do you actually sort of reify that story for everybody so they actually can say, yeah, okay, these 3 people—
Yeah. So, so I, I think you've radicalized me in my story understanding of human behavior. So, uh, like putting it that way, I'm like, yeah, I'm gonna, I'm gonna dive into my humanist roots here and be like, yeah, I don't, I don't know where truth went. I don't know if it ever existed. Um, but it's definitely heading out the door. And what does that leave us with? That leaves us with the cacophony of claims, statements, and narrative structure. And that, that is, that's what we have access to. And gaining access to that, it allows nefarious actors to do things. I mean, look at Grokopedia. What is the point of Grokopedia? Grokopedia, the point of Grokopedia is to influence future language models. And it is a version of the truth. It is a version of reality that has, you know, information and narrative structure into it that is a counter-narrative to that driven by Wikipedia. And this, this is like the world we inhabit, right? Is flooding that information space, knowing that agents are automated, agents are going to be acting within that space. And so the, the power and control is really aimed at that info space. I think, you know, the fact that Musk is working so hard to create a counter-Wikipedia tells you that there must be some intrinsic value of, of story, public story structure. And the goal then, you know, how does Nuremberg trials, how do they happen today? It's like, it's the cacophony. And that, you know, that's the problem. That's the space we inhabit is like that consensus is gone. I think that, that sense of like, independent justice is eroding. I have no idea what, like, society looks like after that, and it's probably not good. But I'm, I'm kind of accepting of it as like a clock or a structure we can't turn back, which means we're left with, you know, really trying to get a handle on that information space where things are coming from. Like, I think they're— that— I think the metadata may ultimately prove to be the most important, which is which— what community, what actors and communities are saying what. And that might get at some of the, the truth problem and the verifiability problem is like, we, we don't know if it's true, but we know this person is responsible for this, or this community is responsible for this piece of information. I don't know if that's satisfying or not, but I mean, it's, you've, you've hit the nail on the head of like a really, really scary, nasty problem.
Thank you very much. Mike Tom has the next question.
Yeah, I actually have a sort of a follow-up to what we just discussed with Max. Or rather, I should want to, to, to step a little bit back for, for myself to better understand your, your, uh, position. Let's, let's concentrate not on a sort of a nasty state of affairs we are currently in, but speak more generally. Okay, you have this cacophony of, of voices, and you, you say, well, basically what you are— what you want is to, to bring some sort of structure which in real time organizes it in some sort of narratives. Can you give maybe some examples or some ideas, uh, more whatever can— more or less concrete, whatever you want, about how these these narratives can be used downstream. So what is, what is, in your opinion, sort of a good practice of, of, of, of using them in, in, of, of actually using them in order to, to, to, I don't know, to, to advance science or to make society more reasonable or whatever?
Mm-hmm. Yeah, I think, I think the, the stakeholder application needs some some, some deeper thought, and, and that is always my weak point. In other words, you know, I'm a very introverted data analyst, and so thinking about how the world uses things is hard for me. I think that's where, you know, Taylor, our collaborator at McGill, is very useful. The way we've conceptualized this is that this is distillable, synthesizable, actionable information for different stakeholders. In other words, if we Oops, sorry about that. I forgot to turn on airplane mode. If we take, you know, I don't know, climate change, the, you know, my version of the narrative of climate change is one kind of thing because I consume certain media, I have certain beliefs, right? So this is, this, you know, the info ecosystem I inhabit has sort of a narrative of climate change. And that is its own push and pull, right? That's constantly shifting and changing. And then there are other communities are gonna have very different examples. And if you are a government that wants to address climate change, then you need to know what those belief structures are, right? And one way is to survey people and ask them questions. Another is to try and infer These narrative structures, and that can give you information that when you're going to do your communications or come up with policies, that you need to align it with those predominant narratives, or you need to create communication that tries to move that narrative in a direction where you want it to go. And so, like I said, this can be used nefariously as well, right? You're you're a government that wants to limit climate action. and promote fossil fuel use. And so you look at what those competing narratives are and you see, okay, well, there's a community out there that has this set of beliefs that these actors are actually not super positively valued. And so let's try and carve them off like classic politics, right? You're like, okay, people have really mixed feelings about environmental organizations. Let's start to tell stories that really showcase how self-interested and And you know, self-dealing these organizations are. And so we're going to move the narrative around environmental organizations from people who do things to help the environment to people who do things to enrich themselves. So you know, there's that problem, which is like we are giving information that is agnostic in how it's used. But the hope is that it is indeed actionable, and that it does facilitate. You know, if we have democratically elected governments. governments, that they are serving the interests of people, and that they would take that information and use it productively. I realize those are big assumptions that are being challenged right now, but that's, that's kind of how I see the uses of this, right? The functionality.
Because anyway, ignorance is not a solution.
What's that?
But I was just sort of making it— ignorance, ignorance is definitely not a solution. To know nothing is, is not Good. So, yes, I mean, of course, AD knowledge can be misused. That's for sure.
Yeah. Yeah. I mean, I think the empirical question would be, does knowledge of narrative, the collective narratives of these wide-scale belief structures that have like narrative structure to them, are they more informationally rich, more valuable than survey data? You know, what is the value added when you ask people a bunch of questions? Questions and the biases you get from that compared to, again, this loud actor, loud voice problem of like, we're now aggregating a lot of like noise. Is that higher value? So, you know, I think there's an intermediate step here of validation, which is like, I imagine this can be useful information because the beauty of language models is you can take all this messy data and you can turn it into, you know, here's the narrative of climate change according to Albertans in Canada versus Quebecers in Canada compared to, you know, country X, whether you can get better information through surveys or more timely, et cetera. Like, I think that is also, that's going to be an empirical test, right? Like, we, we have to also prove that this is actionable information. Right now, we just want to prove we can extract it.
So you are, you mentioned that governments need that information. Shouldn't Shouldn't it be the right of every human being to have a situation of the opinion/belief structure landscape in the same way that we should have the right to see what, I don't know, the weather report is when we fly across the Atlantic or when we board a ship? Isn't that something we should be able to know as everybody, or is that something that should be the privilege? And I mentioned this is another scary question in a sense, because, uh, you know, since 1980, there's like a subset of, uh, the population in large rich country— countries that has privileged access to information for, you know, nominally national security, for example. Um, and also there is private companies who have a lot more information about like what my mom is talking to her friends than I know about that. So is that something which, uh, sort of like, you know, what you're aiming for, is that a kind of better report for like 3 elected guys, whoever they were elected, or is it something Canadians should know?
Yeah, I mean, I, I, I like to think it's, it's publicly valuable. Like, for sure, the aim is not like private reports that we charge a lot of money for that only government officials can access. Like, you know, the goal is openness. And putting this out there. It's— I agree, it feels intuitively like valuable information for people. I don't know. I don't know if it is like the weather. Like, it's, you know, I gotta get to campus before the deluge starts this afternoon. So that's like really actionable today information. I don't know if the narrative map of collective beliefs around any given issue is immediately be as urgent for an individual as it is for an institution. But I totally agree. Like, I think it's really important that people know this, right? And again, in the cacophony, I think we get led astray about believing— whenever they do surveys of like, what do you think about people you don't know? Nobody understands it at all, right? Like, our estimates of what's going on out in the world 2 hops away from our social network are terrible. You know, Republicans estimate that liberals are like whatever they are and vice versa. And so yeah, this gives us like the ideas to unearth and make transparent publicly shared information so that we're understanding what others are understanding. There's a bit of like a meta-awareness going on there that would be really nice, I think. And so that's for sure a goal of this. I think it— my immediate Like, I think what it'll help with is a lot of— I mean, not that, you know, I don't know what journalism's future is, but my hobby horse is the way journalists really struggle with generalization. And so they will constantly talk about the narrative of X, the collective, you know, people believe that, and they're not going on anything empirical or concrete. And so the idea would be to give people in any position where they're making these generalizable claims, about collective beliefs, more tractable, more accurate information.
Thank you very much. So, do we still have some time? Can we ask some questions, or should we—
Let's, let's take a— how about one, one wrap-up question?
Okay. One thing we didn't talk about is your network structures, which, you know, you showed these And they were very generic. And you said like they were produced by a large language model, which was funny because the network structure looked like a large language model in a sense. So do you have any idea of like how that would look like? What do you expect? Like what would be the property of these, whatever knowledge graphy network comes out of there? Also, you're not the first who does this, right? I mean, like if you think about like the Google Knowledge Graph was extracted from the web based on also Google N-grams and Freebase and whatever. So, so there's a sort of like, there's an interesting feedback loop going on, I think.
Yeah, yeah, I, I think, yes, absolutely. I have no claim to innovation on the knowledge graph front here. I think my sole goal is to think about a narrative knowledge graph, right? Or a narratological knowledge graph. And, and what, what units does that consist of? So you, you probably noticed, like, in my elementary definition of narrative, we have time, and space in there too. And I kind of left that off simply because I was doing variable trimming and trying to get to, you know, the most simple possible representation that also contains the most amount of information. And the way I theorize that is, you know, actors are at the center of all narratives. We can show that, you know, some of my research has tried to demonstrate that. That's also just intuitively obvious. NLP loves event structures. I think they're right. And so, this kind of actor-event double pair is something that then we just add a story moral layer to. In other words, actors plus numerous events result in some sort of value that goes with that portrayal and selection.
Right.
And I think as part of that story— so there are different ways to think about story morals. One way is as a general value statement. The other, and this is what I was playing with in terms of making valence explicit and the character role explicit. In other words, we've done work on trying to scale up actors from individual names to roles. So in one, The savvy political strategist is a role that different people can play, and it has a narrative function. And so I think one of the things that's just a little fuzzy right now, and I was hand-waving my way through it because this hasn't been fully operationalized, is where do you— how do the specifics and the generals align? In other words, you can imagine a very dense graph of all actors and all events. And you can imagine, like, higher-level synthesis where you've got all roles and events that have either potentially been summarized, or roles and morals as structures. So I'm gonna just like hunt and say, if you have follow-ups over email, I would love to hear them. I'm also— I just love collaborating. So I, you know, anybody who's like, I want to work on this, let me know. Yeah, so I'm gonna kind of punt on the specifics on it. But that's— those are the problems we're seeing. Like, one of the problems of all of this analysis is scale. In other words, you've got literal words that occur in a news article, and then you've got to get to unified. I mean, this is in Tangerlini's work, like the main pipeline challenge that they had. had 2 or 3 years ago is like, you have all these random names and you've gotta unify them into a single actor. I think LLMs help with that tremendously. But that, like, you know, unique Donald Trump identifier versus unique politician identifier, like the role of politician, then what kind of politician, savvy politician, corrupt politician, right? These are all different valences getting attached to particular roles. in which people occupy. And then when you play a role around certain topics or events, there's gonna be a moral or message, and it isn't clear where the redundancies are. In other words, if you know the role, you know the moral. If you know the actor, how predictable is the moral, etc., etc. So these are the dimensions of this narrative knowledge graph. Needs a lot of work.
Nice.
So as we close up, I have to announce what's happening next. So in 2 weeks we got Laura Nelson, November 17th, and she's from the Centre for Computational Social Science at the University of British Columbia in Vancouver. So in some sense, spiritually, we will stay in Canada. And Andrew, this was great.
And next week we have Christian Benz.
Oh, he's not on the website yet, right?
I don't know why he's not on the website, but—
I don't either. He's on November 12th.
12, exceptionally on Wednesday.
Here he is. Oh, he's further down. I don't know why. Okay. Okay, so this— it's listed in the wrong direction. I'm sorry. So Christian Binz, indeed, it's on the website, but it's not in the right position. He is actually from the University of Saarland in Germany, and we're looking forward to that. So next week, Open Letter. We're looking forward to both of them, but wrong day. It will be on Wednesday.
Thank you very much.
Thanks, everybody.
Thank you.
Thank you.
