Abstract

Ruth Mostern directs the World Historical Gazetteer (WHG), open digital infrastructure that links historical places across languages and time periods through reconciled records and phonetic search, and the Institute for Spatial History Innovation (ISHI) at the University of Pittsburgh. She argues that place is a contested, changing object that gazetteers can model and maps cannot, then describes how the WHG reconciles records against Wikidata and GeoNames, holds 2.2 million places in its public version and about 47 million in a development version, and what funding and staffing keep it running. She also reports an experiment that ran headwords from an 1857 gazetteer through the WHG application programming interface (API) and found the results largely wrong for Asia.

Chapters

Transcript

AI-generated transcript. It was produced from the recording by speech recognition and cleaned up with AI. It has not been edited by the speaker and may contain errors. The captions on YouTube are AI-generated in the same way. Each timestamp links to that moment in the video.

[0:15] Let me just start with some preliminaries. I know it's 1129 and not quite 1130 yet, so I don't want to start the sort of formal talking part of my presentation for another minute. But I know that this is a hybrid conference, so welcome to all of you who are here in person. Welcome to those of you who are on Zoom. Is anyone following on Zoom? How many people are on there? Just so I know who I'm talking to a little bit. But welcome to those of you who are out there. I am not going to be monitoring Discord or Zoom or any other platform while I'm also up here talking or during the Q&A. But if anyone who's here in person is monitoring Zoom or Discord, and when we get to the Q&A part, if you want to sort of raise your hand and tell me what the folks out there remotely might want to bring to this conversation, please do so. I know this is the first session or the first kind of breakout session of this conference. So I've been asked to speak for about 40

[1:18] minutes, 40, 45 minutes, and then that should leave us plenty of time for Q&A. Welcome to Building Historical Infrastructure at Scale, the World Historical Gazetteer. I want to start by talking a little bit about what historical problems we have been trying to solve with the World Historical Gazetteer. One of the other things that Louis wanted to use this session for the purpose of bringing into our collective conversation at this conference is also to talk a little bit about what it takes institutionally to build and maintain a piece of infrastructure that is intended to be useful throughout the digital humanities and throughout the computationally kind of enhanced world of history that we are exploring. here together. So I'll start with those two topics, introduce some of the functionality of the gazetteer as it stands now, and then talk about some of the

[2:19] ways that we're trying to bring more algorithmic AI kinds of methods into the work that we're doing. So the problem that we're trying to address at the World Historical Gazetteer, which has been around for about 10 years is the idea that space, that place is complicated and multivalent. And place is really, really important to people. People are literally living and dying over whether they get to be in places that are important to them and whether the places that are important to them are allowed to have the names that they care about and that they associate their ancestral dwelling and belonging with. And in a very frequently used definition by the geographer John Agnew, what it means to say that some place on the surface of the earth is a place is that it exists somewhere. It has a locale. And what Agnew means by this is that it has buildings, infrastructure,

[3:23] stuff created around it, names, etc. Right. So we get to say that Johns Hopkins is a college campus, right, because we kind of know what college campuses are like. And because this place has the kinds of infrastructure, buildings, senses of meaning that a college campus has. So that's what he means by a locale, right. Johns Hopkins University is a locale. It It has a location and it's a site of meaning, right? Everybody who's here, including those of us who are here just temporarily, just for the conference, right, imbue this location on the Earth's surface with some kind of memory, relationships. We network it into other places that we know about. So that's really, really hard to represent. And maps, right, paper maps, right, the things that we think of as being the sort of kind classical tools for spatial representation don't necessarily capture all of that. For those of us who are historians as well as or rather than geographers, we also

[4:25] know that all of these things change over time, right? And all of these places change their meaning, even literally their locations, certainly their names, very, very much so the ways in which they are sites of meaning over time through what I like to call placemaking events, right? And just a few examples here, right, of literally a farmer moving the location of the French border because he found a stone that he didn't like while he was plowing his land, and it turned out to be an official border stone marking the French -Belgian border. Of course, contestation over places in historical Palestine, right, about whether they should be referred to by Arabic or Israeli names. And then also, right, event from Black Lives Matter, pulling down a statue in Richmond, right, what does it mean that that plinth holds one kind of meaning or another? And so all of these are examples of what I think of as placemaking events relative to this John Agnew definition. So in building the world historical gazetteer.

[5:27] The question is, how do we not just have a map, not even just have one static database, and instead have an information system that allows this kind of contestation and these kinds of placemaking events to be brought together, right? Another example of what I mean by placemaking events, this is the Wikipedia entry for the place names for the names of Beijing, right? Beijing has been known by many, many different names over time. Each of those name that you can see are ones that have occurred in the context of a regime change, right? And so thinking of political history, not just as being about a new regime taking power, but also about being reflected in these kinds of placemaking events, right? And the sort of implicit storytelling behind each of these name changes, not just saying Beijing has had many names, but saying that those names are meaningful, right? There is a historical narrative that unfolds behind each of those names. Another example of the same thing, one of my favorite projects, still

[6:31] up on Flickr, which still exists. I am, right, every time I check this site, I'm like, it's still there. Someone has to, like, index and archive this site, one of my very, very favorite sites, which basically is somebody sort of scraped a bunch of photographs of places and then tagged them. This is many years ago. I don't actually know what technology it was. 2010, I don't know how this work was done, but basically if somebody uploaded a photograph and all of their photographs for a month or more were from that same place, they got tagged as a local. If the photographs from that place were not from the place that the rest of that person's photographs were from, they got tagged as a tourist. And what you can see is that the tourist version of what it means to say, Horton in New York or Tokyo, Beijing, Singapore, anywhere. Right. And there are dozens of these places. It's such a cool project. You should all look at it. Right. The tourist version of what makes a place a place is totally different, wildly different in some cases than the local version of

[7:36] what makes a place a place. So these are just sort of different examples of unfolding the notion of why is place a complicated, interesting sort of object of information that deserves more attention than you would get from just like a point on a map. One more example. I think this is my last one that I love. I can go on. No, it's not my last one. I can go on and on about this topic. Right. Really, really cool example from a couple of years ago, similar to what I just said, but from the from New York Times reporters fanning out and asking New Yorkers, what neighborhood are you in? Right. And getting some answers that are really clear, others that are absolutely contested. And the reason I want to put this particular slide up here is for the purpose of saying this isn't just like, oh, cool, look, there's diverse knowledge in the world. that this actually has really, really meaningful implications for people's belonging, ability to stay in their home, ability to afford their

[8:39] home, right? This is, we all know about gentrification. This is gentrification as something that unfolds, among other things, through processes of place naming, right? And so it's the kind of thing that we would really like to be able to model in an information system. Also, possibilities for counter mapping and resistance, right? Being able process of place naming and of linking lived experience, including resistant and minoritized lived experience to specific places on the map. So many really good examples of this in the world as well. Okay, so what this means, so pivoting, and I don't have a good pivot transition slide, so listening to my human voice, we are now pivoting right to the next topic, which is the genre of the gazetteer. So a map, we know what maps are. Maps are visualizations of spatial information. Gazetteer is basically an unusual but really, really useful word for a

[9:41] data set, a database of some kind, even just a list of place names, right? Even the index in the back of a paper atlas is a kind of a simple gazetteer. And so one of the cases I want to make, whatever projects you all are all working on, whether or not you're interested in using the World Historical Gazetteer that I'll be talking about in a couple of minutes, making the case that for us as historians, in many, many cases, maps are not particularly useful. Maps are cool. I love maps. These are two beautiful maps. I love this Mississippi map. If you don't all know, it's one of the best maps in the world. Maps are really cool. But in terms of being able to model this notion of placemaking events. You can do that really easily with a gazetteer, pretty simple data model, right? I can talk more about data modeling later if anyone is interested, but basically you can have a list of place entities. You can have a list of place names. You can have a list of placemaking events. You can wire them up. And therefore, instead of just having one line that says,

[10:44] this is the Mississippi River, you can say, there's one entity. It's still the Mississippi River, but it's occupied many, many different channels over time, some simultaneously, et cetera. Same thing on the right-hand side. We have Poland, right? There's an entity of Poland that Poles would argue correctly. It goes back hundreds and hundreds of years, thousands of years, right? But it has occupied many, many, many different kind of spots on the Earth's surface over time. It's had many names over time. It's been conquered. It's been an empire, right? And you can represent that in a gazetteer with entities, placemaking events, and other attributes like spatial extent, right? So gazetteers are really, really, they are the genre for doing that. Right? Gazetteers can do things that maps can't. I think we've talked plenty about that already, Right. Interesting patterns might not be geographical. Geographic data might be more useful to visualize in a way other than a map, for

[11:49] instance, in the form of a timeline. Geographic information might be uncertain, ambiguous, missing, change over time. Right. Showcasing images and documents, showcasing contestation. These are all things that are really hard to do on maps, but they are what gazetteers are purpose built for. And then sort of just all of what I've just said for the last 10 minutes is kind of summarized in this programming historian article that a colleague and I wrote a couple of years ago. So any of this information about what is place, how to represent place in a gazetteer, why you might want to do that, how to link gazetteers together, this is a place you might start for that information if you're interested. So transitioning now to the World Historical Gazetteer. So the World Historical Gazetteer, which is what I'll be talking about for the rest of my time with you, is a way to take multiple gazetteers

[12:50] and link them together. Right. So as I started out saying the problem that I really want to solve and that I think is important to solve, both from an information management point of view and from a sort of political and ethical kind of point of view as well, is the fact that a place might be known in many different ways by many different names. And what that means is that any individual data set is unlikely to ever be able to capture the range of information that we know about any given place. And the solution to that comes with linked data methodology and building a platform that allows anyone who has a gazetteer to contribute it to a shared system and then to use, in our case, a sort of SCOSC, well, in our case, in the present, but hang on, I'll be getting to AI, right? Basically, to be able to say, my place that I call Beijing

[13:52] and your place that you call Beijing actually refer to the same entity on the Earth's surface. And that's something for which we have built the World Historical Gazetteer. 2.2 million place entities in the public version, 47 million in our dev version. And I'll talk about basically it's sort of using kind of AI automated linking algorithms is what's sort of allowing us, we hope, in progress to get from 2.2 million making semantic sense to scaling up very rapidly to 47 million. So it's an index of all these place entities. It's a tool for publishing information about these entities, for enhancing your own gazetteer with information we already have, for creating links. And then it's also a builder for annotated collections of information about places. So World Historical Gazetteer was founded about 10 years or nine years ago, I think

[14:53] is when we got our first grant and started working on this. We went live maybe seven years ago or so. And for the first, you know, until last year, this was always had some built in kind of linking algorithms, but was always very, very absolutely human in the loop and human scaled. So I want to talk first next about what we have been doing sort of at the kind scale, and then what we are now experimenting with what we can do if we sort of scale up to the amount of data beyond which humans can sort of meaningfully be involved in every single step. So, oh, and then also, right, one of the things Louis, I think, wanted me to talk about in my talk, thank you for being here at this session, is what does it take just institutionally to do all of this, right? I think it's pretty frequent in these kinds of conversations, especially for a

[15:55] relatively mature project like this one, for people to get up and say, look at this amazing thing we built. It's mature. It's up there. It's online. You can use it, right, without talking so much about sort of how it's anchored in institutional relationships. so the World Historical Gazetteer exists as a project as sort of an institute that the University of Pittsburgh my home institution was willing to fund at a certain point when I was able to persuade them that I had done something interesting and big and useful so it's sort of the Institute for Spatial History Innovation is a kind of a wrapper around the World Historical Gazetteer but is also a project for fostering other kinds of projects in Um, there's a link to it, a URL for it online if you want to explore it. Um, and you know, we couldn't do this if we didn't have, um, some staff. And so I want to acknowledge my collaborators, the people I work with. I keep saying we, and we basically means the people who you can see there at the bottom right. So, uh, myself, uh, my technical director, software developer,

[16:59] Stephen Gadd, uh, my graduate student assistant, Yating Leo, um, Alexandra Straub, Palak Vashist are the two sort of project managers and editorial managers at ISHI. So this is really, in the scheme of digital humanities projects, something that is pretty infrastructural, but also therefore needs an infrastructural amount of funding and staffing. And I think that's something that I hope over the course of this conference, we're going to be inquiring not just about sort of the cool things that we can do technologically and some of the challenges and affordances of the cool things we but also what does it actually take to make something like this actually function? And so I just want to prompt some of that conversation that we can pursue in the Q &A if anyone's interested. We've had a series of grants. We've had hard money funding. And we're existing now by the grace of my university on hard money funding that's keeping this together, as well as also really pretty significant comp server

[18:05] space. So that's one of the things that a project of this scale requires. Okay, so what do we do at the World Historical Gazetteer? And so, as I said, it's a project. Basically, it's a platform that exists. It is up and running. You can follow along, whgazetteer.org. I should put the URL up here, right, that you can search our index. You can contribute your own. I mean, basically, these are the things that you can do, right? You can query the index and find information about the 2.2 million place entities that we know about in the public version. You can build your own collections of places of which that includes both data sets and then sort of annotated publications about places that are linked thematically. We have teaching tools and documentation and announcements. So, okay, so what happens if you search for a place? So this is very small on this screen. Like, what place do I have here?

[19:08] Right. Oh, yeah. And this talk is also sort of an archaeology of all of the past places I have done versions of this talk. So you got some Poland before. Here you were getting the city of Lublin, where I did a version of this presentation in May. Right. So you search for Lublin. And what you learn is a couple of things. One is that there are quite a number of different places called Lublin, right? And so you get a list of all the possible Lublins as well. And here I'm kind of prefiguring some of what we're doing experimentally with more algorithmic methods. So we've built a sort of phonetic finder. So I search for Lublin and this totally automated phonetic finder, which is working okay, kind of thinks that maybe I'm interested in Lublin. There are seven linked records for Lublin, but maybe I'm interested in Ljubljana or Lubin or Lubin or Lubai, right?

[20:08] Just like here's some other stuff that has some phonetic similarity, right? I could constrain that a little bit. I could put on a spatial filter. I could filter it by the types of places I'm looking for. And these are the Getty Thesaurus of Geographical Names, kind of high-level place type categories. But totally unfiltered with any of this, it's just trying to do some phonetic matching. But the Lublin I'm interested in is likely to be the one that shows up first, and it's likely to be the one that has seven different linked records attached to it. So if you look at the linked records, and here's what I'm saying about, about. And of course, like all of, you know, I can talk about the sort of what we're trying to do theoretically and what we've been able to implement. Of course, there's always a little bit of a gap, maybe even a big gap. But what you can see is that we have a range of different data sets, each of those sort of cards that you see on the landing page on the right, right? So this is a PID. So it's a permanent ID for the concept of

[21:09] Lublin, right? And what you can see possible names from many different possible data sets, even many different possible point locations. They cluster together, right? But anyone in their own data set can say, you know what? I think the best possible representative point for this city is here. Someone else can say it's here. But as long as they establish that there's a close match between the two, that's enough to say, okay, I want my data set to show up in this cluster of linked data sets about this sort of conceptual entity called Lublin. It's not its name. It's not its location. It's this conceptual entity in that sense that I presented at the beginning of the talk. And, right, and so what does this look like? So then another example from Bamberg. I've been hanging out in Central Europe a lot recently, right? Same thing. You search for Bamberg, all the possible records for Bamberg,

[22:10] right? You learn, we have this sort of, you know, automatically generated information about its geographical setting, its altitude. We know, you get the sort of summarized information up here that we have 17 distinct names. We get information about the temporal extent that we know about from this range of six sources. You get a little bit of spatial information. From there, you can write collections of records submitted by data set owners, and then you can explore any of those particular places. If there's one specific, you can decide, you know what, of all these six linked records, the one I'm most interested in, the one I trust the most, the one I'm most curious to explore is the one from the old world trade data set. And so you can get back specifically to that data set. were to have its own URI with its own independent landing page, you can get directly back to the actual landing page with whatever additional contextual

[23:12] information there is, citations to primary sources, images, rich sort of qualitative information about that place. And so that's the way that we're trying to do this sort of range of data linking. And so this is, again, And this is all human work, right? This is humans creating data sets, and this is humans by hand in a sort of computationally aided system that I think I have a slide about, attesting by hand that their place is a close match to a place that's already in our index. We also have a layer for ecological context, so you can turn on this information and find what biomes any place is in, what watershed they're in. And so that's one of the other things we think is cool and important. We have, I think it's now up to more than 100 data sets since I last made this slide. But in addition to searching for any place and finding all those linked

[24:16] records, you can also browse data sets and find any of the data sets that we know about. and then what does this look like, right? And then from the publication, right, say you're interested in this, this is more Poland, right? So this is 16th century Polish provinces. You can browse the data set. You can go to the data set, find more information, read the metadata, of course, and even if you want, you can download this and you can go back to the original site where the information was published with its fuller context. So this is really the way we're trying to instantiate this linked data methodology. And again, with the idea of always being able to illustrate multiple and contexted meanings. And so this is one of the data sets I'm really proud to have fostered. When was this? In 2020 or so, when the new Secretary of the Interior here in the United States determined that hundreds of

[25:20] names that were still officially existing names for United States places that used a very offensive name for indigenous women would no longer be the official names. And so we created a data set of all of the replacement names. And so you can see here the overview map of where those offensive names had been located and then what the new names were. This is also kind of a point of limitation in what we've been able to implement so far. And one of the things we're actively working on is that right now, we don't actually have sort of name change types in a kind of authored algorithm, right? And so if you were to search for any of the names that the offensive names were replaced with, you would get a landing page like, Like this one, right, that includes both the replacement name and the offensive prior name and doesn't actually give any guidance to searchers

[26:22] about what happened or why they might prefer one name to another. So that's kind of an open area, like we are aware of the problem, and it's an open, I think we're aware of the problem and the solution, but we have to implement this. And so this is another thing we're figuring out whether through human or algorithmic methods, how do we tackle that problem is an open question for us. And then same thing, right, again, historical Palestine, same kinds of issues, same kinds of situations of dealing with deeply, deeply contested information about how names of places have been replaced with other names. We also have collections. So these are, so in addition to data sets, we also, and again, hand authored, I'd be really interested in fostering some conversation about what it would look to scale this up algorithmically. But these are totally human scale publications where and and this is teacher facing. So, for instance, right, we have one about like the K-pop band BTS that a

[27:24] teacher brought together in order to explain to his students why there might be individuals, talented young men born all over Korea who all moved to Seoul together, all signed fashion contracts in Paris together. Right. If you come from rural Korea, that is not where you're going to make your career. And so he created a really cool kind of publication and lesson plan about that. Right. But this is our place collection feature where we can now see this is another really cool project that one of my undergraduate student workers did on the origins of the Bulgars as an ethnonym and an ethnic group in medieval Eastern Europe. And now if you search for any of the places, right, if you search for Byzantine, Constantinople, Istanbul, of course, a place with all those names, you learn about all of those names that we know about from all these different data sets. But also you learn that it's connected thematically to other places that you might not necessarily associate it with as a result of being

[28:29] part of this historical narrative of the sort of establishment and spread of the Bulgar people. So that's our place collection feature. So, okay, so pivoting now. No, this is, I'm not pivoting quite yet. We also have collections of multiple data sets into a data set collection where if there's a larger project that wants to bring together multiple big data sets and say, okay, all of these collectively represent some complex and interesting phenomenon, there's a landing page, there's a linked essay. This is from an experiment we did in-house about what is Central Asia even, right? Thinking about Central Asia as a region, right? And recognizing that that's something that's very complex, very contested. It takes multiple different data sets to be able to describe. You know, it's one thing to say you can put a point or a polygon on the Johns Hopkins campus. It's another thing to do something like that for the vague concept

[29:29] of Central Asia, which certainly is spatial. Maybe it's a place. A really interesting conversation to have about what it would take to scale this up in a more machine-aided way. And that's one of the things I would love to talk about in the Q&A. Built on linked open data principles, I've talked about that already. So we have lots of good documentation both on our site and on GitHub. I don't think I want to talk too much about that here, except to say that the idea of linked data, I guess I've already said this, is about sort of establishing until now. And I mean, actually, one thing I realized as I'm talking, I haven't really explored is what happens to linked data in the age of AI, right? That is, linked data, a sort of web kind of 2.0, so 15-,

[30:33] 20-year-old concept, is about establishing sort of hand-coded networks of information that define phenomena, therefore takes a huge amount of human judgment and human time. And what would it take to create those sort of networks of interconnected information in a way that's more algorithmic, but still sort of allows for the kind of human judgment, walk back to sources, right, and information and metadata that explains why sort of clusters of linked information have arisen in the way that they have. Again, I think open question for my project and probably for us all here at this conference. Right. So we do have tools for creating data sets and data set collections. This is all up on the website. I'm not going to talk too much about that now except for the purpose of saying something, just kind of reiterating what I've said

[31:37] already, that until now, this has been a human-aided process. A sort of computationally aided human process, where in this case, for instance, somebody has the target place of Babylonia from their data set, and they get a couple of suggestions from our sort of pre-AI algorithmic system, I guess you might say, of things that we already know about from Wikidata and then from our own index that may or may not be. be the same as the target place that the author of this data set has contributed, and they have the option of either saying, yes, this is a close match, or no, this is not a close match. I think there's a lot more work to do to make that authorship decision more explicit. On the one hand, to make authorship decisions like that more explicit. On the other hand, to automate at least the simple cases so that authors

[32:38] don't have to walk through this for every single entity. And so I think this is a case where what AI offers us is the chance to both to automate things that can be automated, but if this makes sense at the same time, therefore to be able to focus on and isolate the kinds of things that still need humans, right? So that humans make human judgments and can focus their time on the interesting, messy, complex, ambiguous cases of judgment and let the machine do what the machine is good for. Right. And I think saying that this sort of AI moment is an invitation in both of those directions is really, really, really important. And one of the things that I hope we can sort of think through at this conference and beyond. Accessioning is the way we do it now. Two step process sort of first you reconcile against Wikidata and GeoNames and then

[33:41] you accession into our WHG index with all of the AI. other information that we've published from specialist data sets. Preview of the new version. So as I said, this is where we're really starting to think hard about what it would take to scale up from the sort of single digit millions of places, which is still plenty. That's a lot of place entities, right? But also at the same time is a tiny fraction of all the places in the world, Right. I like to say it's either a lot of information or not very much, depending on how you think about it. And so behind the scenes in our dev version, we have about 50 million places that comes primarily from indexing all of Wikidata and all of GeoNames. Right. This is also part of conversations. Katie McDonough is in the room. No, she's presenting at the same time. She's not in the room. Right. But also from MapReader, right?

[34:43] Machines reading maps from the, you know, tens of millions of names that her project has been able to sort of scrape off and georeference from historical maps, the entire David Rumsey archive. Right. We have been talking for years, but I'm like, we can't we can't deal with tens of millions of names. Right. Our platform does not permit that. Our methodology doesn't permit that. essentially is using semantic clustering to replace, I think more appropriately to enhance that hand reconciliation process that I presented already. And so this is absolutely, now we are in the world, so finally as this talk is proceeding, of this sort of algorithmic world where we are taking some of what we used to be able to do with human judgment but couldn't do at the scale we and instead scale it up to work by a computer.

[35:45] What this means in the way that we've implemented it so far, I don't think I have exactly the slide I wish I did here, is that as a user, you search for a place. Optionally, you set what region of the world you're interested in. You can tune it. We have some sliders saying definitely match on the text string. definitely match hard on a specific location, match on a feature type, right? All of those are sliders where you can set larger or greater sensitivity. And then you get a sort of automatically generated cluster. So rather than a hard landing page that says, Babylonia, here's a list of eight records about it. Beijing, here's a list of 10 records about it. Instead, it says, well, you know, based on the criteria that you've set, how much sensitivity you want, here's a bunch of things that maybe kind of sort of are Beijing, right? I think we obviously gain a lot from that.

[36:47] Potentially, there's also a loss for that. And what we're now working on is what a hybrid system, right? For instance, thinking about the offensive names for indigenous women case, right? We would not, I don't think we want to lose the fact that those two names that are entirely unrelated text strings have a really, really important relationship to each other, right? Like that's the work of history, right? That's the work of being able to say, as I did at the beginning of the talk, that place is not just about, I don't know, a reference work. It's also a way that our historical narratives of meaning and belonging and unbelonging and conflict are instantiated. And I think that looks like some hybrid system that we haven't quite developed yet. So again, something I really want to prompt some conversation about. And so this reconciliation API that we're now using is phonetically backed.

[37:50] But as I think I showed you on a previous example, the phonetic system is only as good as it is, right? And then, right, links out to contributor projects, which we are doing already. Okay, last thing I want to talk about, and being mindful of the time, another way that we are really doing some serious experiments with AI is whether we can use named entity recognition, a sort of kind of super-powered, AI-powered named entity recognition in order to bring more rich, interesting, numerous historical places into our index. Again, sort of how do we move from sort of by hand creating gazetteer datasets to being able to create gazetteer datasets that are good, cool, really large, based on print information and bringing them into our system. So this is an experiment we did pretty quickly

[38:50] So we started with this gazetteer of the world in seven volumes. I want to give you a sense of this as a sort of tactile material thing in the world. It's 1857. We are in the high Victorian era, and we have some British geographers who make this giant tome, this encyclopedia with their 1857 spellings of all the places in the world that they think exist. Really cool, interesting source, both from the point of view of the history of imperialism, right, the history of knowledge, as well as also how interesting would it be to bring these interesting alternate spellings of places around the world as they existed in 1857 into the world historical gazetteer, right? And this is already usefully for us. This is already, it's on Internet Archive. Great. It's out of copyright, of course. It's been OCR'd. Awesome. So we are good to go, right? So this means that we can use some agents to crawl the book really

[39:52] quickly. This was like a one-day job over the summer to find all of the head words, to run, right, the head words meaning like, right, the names of the places, to run them against our index, we're done they're on a map right but um most right and then more and then for those who are interested in the technical details which again for reasons of time and honestly my expertise this was done mostly by our developer i'm not going to get into the technical details but for those who want it is all up on the whg blog so you can take a look at those details um right so um And again, what were we able to do? Thinking about the World Historical Gazetteer not just as a growing resource, but also as a tool. Welcome to all of you. It exists. The API is out there. Please, please bring this into any of the work that all of you are doing. We used our API to take that sort of raw list of

[40:53] headwords from the 1857 Gazetteer and run them against our existing index. This is something that's free and open for anybody to use. Almost all this information is wrong, at least for Asia, which is, you know, what I wanted to stress test because it's my area of expertise and also because that seemed like where the system was most likely to fall down. It is, in fact, very wrong. So and again, right, this slide deck is an archive of where I presented this work. I last presented this in Incheon, Korea. So you have an example from Incheon, Korea, except that I searched for Incheon. I found this Keim-Khotun, right, which is in Chinese Tartary, right, and yet is located on the map just west of the city of Seoul. And so this is basically where essentially the problem is that our phonetic search tool was failing.

[41:56] It is just not very good. And we've now sort of fine-tuned it a little bit, but even so, it is nowhere close to being feasible to actually use for implementation with confidence for a project like this. You know, what I'm trying to do in this talk is both prompt what we can do at scale, what I think some of the promise and possibilities are for working at scale, but also the fact that we're really not quite ready, I think, with a lot of these tools. And I suspect that's going to be a theme of our time together today and tomorrow. Right. And then, again, on this blog post, more information about what this is, what it isn't. It's an experiment. So there's more information here about what failed about it. And then finally, coming to my last slide, and right, Louis just walked out, which is either good or bad. So I'm so excited about this conference.

[42:59] This is so great. It was the fact that it was able to be planned and implemented and run with as little human staffing as it has is remarkable to me. But this is the last slide that I added yesterday as I was finishing this slide deck and getting ready to leave home. So I got, so on Tuesday, yesterday, two days ago, right at 9 p.m., I got a, we're running this conference with a minimum of staff and a lot of automation, blah, blah, blah. Here's some options for you presenters. And I said, I do not want to monitor my own Discord email or Zoom chat during my session. But if there's a moderator or participant who can do that, that would be awesome. Okay. Then what happens? Right. 10.59. Ten minutes later. Oh, no. One minute later, you're all set. During your session, we will email you new questions from the Discord thread every few minutes. To which I responded, right? Thanks. I do not want to do that. Right.

[44:01] And then number five, 11, 10 a.m., you're all set during your session. We'll email you new questions from the Discord every few minutes. So this is my last slide. This is just sort of my last slide here is kind of just for fun, but also and also any of us, I'm sure, in so much customer service type interaction in our lives. This is very, very familiar. I'm sure we have all had experiences like this. It's here both sort of just for fun. Like, this is what happened to me yesterday, 24 hours ago. But also sort of with seriousness, so very, very much promise for all of these tools that all of us are exploring. And yet also at this minute, still snapshot of October 2026, I think still some really serious limitations that we all need to be reckoning with. Thank you so much. So happy to take questions, comments, conversation.

[45:03] Yes, Marnie. More thoughts, and I have a live question. A lot of this is really contestable information. Some states might choose to rename entities by bodies of water. Just for example, right. How do you manage denial of service attacks? So cybersecurity, really basic question there, because the material you're dealing with is sensitive and unpleasant for some groups. Right. Yeah. No, remarkably, we have not actually encountered either like heavy bot traffic that has made our traffic impossibly slow. We haven't had DDoS attacks. I'm not sure why not. You know, anytime I talk to our developer and I'm like, do we not have a problem somehow when everyone else does? I sort of get a, you know, like he's like, I guess not.

[46:04] So I'm being a little flippant. I mean, I think he's built in some tools, but I don't know, you know, like as you were saying in your keynote, right? I don't know why our box is safe for now. And certainly this is something that keeps me up at night. Yeah. First of all, the online people have asked you to repeat your questions. Yeah, sure. So the question is the last question, sorry, was what are we doing about denial of service attacks, bot swarms, et cetera? And so my answer was remarkably not an issue yet. I've got an actual question. Tell me, what do you see your relationship with Wikidata, which has sort of created a similar but less intentional version of what you've created very intentionally over decades of time? Right. So the question for folks remotely is what's our relationship with Wikidata? So right now, so our data is structured data,

[47:10] right? So I guess so is Wikidata's. But our goal, which is different from Wikidata, is to sort of intentionally keep every attestation of information about a place separate. That is not to amalgamate everything into one master entry, but instead to say, look, you might take this place entity and think of it by this name. And here's somebody who's made that contribution. And here's why they think that name was a name that people were using at a certain time. Here's a totally other data set, right? This data set maybe uses the same name but has a different representative point location for the same place. So the idea intentionally is to maintain that sort of multivocality as one of the core characteristics of how we think of place.

[48:11] Sort of in this, that's how we're sort of instantiating that cultural geography literature where I started out. So Wikidata is such an important source, but it's just one of the sources that we're indexing. The other thing that I'll say, and this is the topic of an application I'm currently, a grant application I'm currently in the middle of writing, right now we have a one-way relationship with Wikidata. So we are populating, we're using Wikidata to populate our index, but we're not sort of repopulating Wikidata with our information. And it makes a ton, like I would like to have, I don't think it's that hard, right, to have like a world historical gazetteer, PID, be one of the things that people find about a place on Wikidata, which is to say, I guess, another way of saying, of course, we are not in competition. It's to very closely with closely

[49:16] related aims that are, I think, doing things that are slightly different, but that should, of course, be kind of mutually populating one another. Yep. It's just handy, so I apologize. I have had subscriptions to Ancestry of MindWare Pass. They have directories, census, newspaper, which are place-centric. Has there been any inter-leaving or work between this and those really huge deposit? Yeah. Yeah. So this is a question for the folks online about the relationship between the WHG and the genealogy community. So the gazetteer information on the big

[50:20] genealogy websites is really not very good. They don't have all of these alternative place names wired up into a place name index in spite of the fact that they have hundreds of years of information from all over the world about people's birth and death and, you know, marriage and life change information replete with millions and millions of place names in all of the different scripts and spellings in which they might appear. So absolutely, I think probably the best global information about place anywhere in the world, sort of digitized yet unstructured, is out there on all those genealogy websites. We're in conversation. So there's a project, a kind of new platform called Tree Alive, which is building spatial tools and other

[51:24] structured data tools on top of the big genealogy websites. And we're in conversation with them about them basically wiring our API into their platform, which would be just, you know, so useful for genealogists. And then I think the hope, again, as I just said with Wikidata, would be to make that a circular relationship. So if somebody finds some, you know, birth record from, you know, the past that has a information about a place that we don't know about or an alternate spelling of a place that we can then sort of repopulate our index with that information. But yeah, I very much have the genealogy community in my sidelines. So thanks. Yeah, over here. So my question is going to be about bias in data, but a little bit of context first. I'm taking a class on digital mapping right now, and we just

[52:25] covered decolonizing data. So I'm thinking a lot about doing so much and being so attentive to all these alternate place names. But that is going to be really hard with as you scale up, of course, because so much of the data sets that you could be scraping are to come from big institutions. So I'm just curious as I'm grappling with the question, how are you guys thinking about that as you're scaling up? It's really hard. Right. Yeah. Thanks for that. So this is the question about how to deal with colonized information and how to make sure that as we scale up, we're not sort of scaling up into only information about from dominant institutions. Right. As Marnie, this is like your World War Two problem. Right. How do you make sure that everything isn't World War Two? Right. Right. And I think this is why and this is something I mean, this literally keeps me up at night. And, you know, I think this is where no matter how

[53:26] much we scale up, no matter how much we're trying to bring in information that comes to us at the millions, tens of millions, you know, hundreds of millions of entities at a time. It is going to be absolutely essential. And, you know, for me as the director of this project, absolutely non-negotiable, that we also are always bringing in specialist data sets. That is data sets created by individuals and that we always have tools that allow those individuals to establish, you know, hard coded authored relationships that say, hey, you know what? But this place that has the colonized name of X, right, also once had an indigenous name of Y. And if you search for X, you must find Y along with it, right? No matter how much we're doing things algorithmically, we don't let you come to a landing page for a place and finding colonized name X without also

[54:30] finding, you know, indigenous name Y. I think that's that's and so, I mean, this is this is a problem that we have to sort of build out better, like with the, you know, offensive indigenous name example that I that I provided, you know, during my talk. I think we need to do a better job in the project of making sure that we're authoring those relationships. That is, we're saying the relationship of these two names is that one is colonial and the other is indigenous. Right. We need to elevate that better. But then also it's the problem that, you know, is kind of the ethics of the whole profession. Right. Which is that, you know, first colonized information is less well archived, less available, has less good metadata, is less likely to be digitized. Right. Is less likely to have the attention of historians. And so some of this is a digital humanities problem, but

[55:34] some of it is sort of the issues at the wider scale of the whole profession, right, and the whole society, which is, you know, how do we make sure that as we sort of race ahead into, you know, more and more aggregational, if that's a word, tools, right, that we're not there for just replicating. and reinforcing problems of data injustice and social injustice that those tools are sitting on top of. Yep. About this question of how do you get smaller data sets, or very hyper-local, what we might call sort of exquisite data sets that are not particularly useful to a vast majority of people, they're not particularly relevant, but really interesting about a very particular place. And while you were talking, I pulled up the guest here, and I was curious because I'm moving to Philadelphia. The house that I'm going to move into says that it was built in 1752. I'm not sure that's accurate. So as a historian, I want to figure it out.

[56:36] So I was playing around to see what your tool could help and figure that out. And it occurred to me that, you know, there are going to be a not insignificant number of people like me with a historical interest who will do the work of going to the city archives, getting some information. And I will build, because I'm a nerd, the kind of documentary record particular place and it will only be useful for me but the magic about what you're building is that in the aggregate a lot of us doing those kinds of things could illustrate a lot of interesting patterns and i'm trying to think what would induce me to do this uh and to share it with you and to like build the sort of thing in a way that's helped for you i thought went to reddit right interesting karma where like posters who are doing interesting looks consistent with the culture of the platform get uploaded. And I'm curious if you thought about creating a gamification element to encourage people to be helpful squirrels, finding acorns. Right.

[57:37] Yeah, no, that is a super, super interesting and good suggestion. And then for the folks online, the question essentially to summarize is how do we incentivize people to contribute, especially people who like nerd out on one individual place, right? And one of the things that's at the very top of my development agenda, and I think it's built into the dev version of the site, is to make it much easier for people to contribute one place, right? Exactly your case. And I encounter this so much when I talk about this project or present this project is like, and this goes with the ancestry question too, I think. I do not care about thousands of I care passionately about one place. I know everything about that one place. I want to contribute to you that one place. And I think, you know, I think this also goes along with the questions about how to be better at our work of decolonizing the project is to make it just dead easy for people to contribute one place.

[58:41] And then I really like the idea. Also, we have like a comment. You know, you can talk back when you end up on a landing page. a correction, but we haven't really showcased it on the site that much. And I love the idea of something like upvoting like on Reddit as, or even if it's a more, again, algorithmic, automated, you know, one thing that we've talked about on our team is making sure that if we have information, say, from Wikidata and GeoNames, and also we have information contributed by an individual, right, a scholar, someone from a descendant community, et cetera, that's always what shows up on top, right? And so I think there's a lot more we can do along those lines. Yep. Fine. Aspen, do you have concerns about the commercial entities sort of taking this data, so she names the big GIS company as free, Mm-hmm.

[59:41] So they might weaponize it or just money might sort of drive the agenda. Yeah, yeah. So the question is about what do we do so we don't get taken over? Essentially, we don't get our content stolen by commercial entities. So all of our – we've just recently been doing a licensing audit. and all of our data sets, all of our contributed data sets are Creative Commons licensed. So if Esri wants to use this, up to a certain point they can, right, I suppose, but they would have to be doing it in a transformative enough way that it doesn't run into our Creative Commons license. We're also open code, but again, we have the license on our code that it can't be used or modified for commercial purposes. So we're actually in the middle right now of

[1:00:42] conversations sort of with the commercialization office at the University of Pittsburgh where I'm based about, you know, like I am not closed to the idea of commercialization. something like Esri or Ancestry.com, but it would be a matter of metering calls to our API. That is, the data itself is always, and by relationship with our contributors, must always be open. But certainly, and I guess in a way this goes along with the DDoS question also, if people, if an entity like Ancestry or Esri is like, you know, hitting on our server in a big they should be paying for that. So that's basically the answer to that question. I think we're at time. So thank you all so much for coming.