Abstract

Atiba Pertilla describes how the German Historical Institute Washington built Migrant Connections, a collection of German immigrant letters that volunteers transcribe on Omeka Classic with the Scripto interface, and how the spinoff project Let the Past Speak combines text-to-speech (TTS) and retrieval-based voice conversion (RVC) to reconstruct 19th century northern Hessian German and Texas German. He then shows how Google Document AI and Claude helped him extract and restructure Baltimore city directory data, including house number changes after the 1887 renumbering, for his study of immigrant financial practices. He argues that AI favors the historian at a workbench over the large digital humanities lab, and the audience discusses voice ethics and errors in archival metadata.

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.

[0:06] All right, so I think maybe I might as well get started. My name's Atiba Pertilla and I'm a research fellow and digital editor at the German Historical Institute in Washington. If you're not familiar with the GHI, we're an advanced center for historical study located in Washington, D.C., and also with an office of the in Berkeley and were funded by the German government. We pursue a variety of projects in the fields of American history, German history, and global history. Today, I will be speaking about a couple of projects in the history of migration space, so to speak. I want to thank everyone at Johns Hopkins for putting this really amazing conference And again, my brief is to talk about migration history and AI, so I'll do my best to give an overview of a couple of the projects I've been involved in, and maybe

[1:10] then at the end of my talk, give some thoughts about the scale of what's going on in history and how history and AI link together today in the current period. So there was someone who I think was going to help with the discord, but I just want to, so I don't know yet whether or not people are listening or able to follow along, but I'm just going to continue for now. So yeah, so I'm going to have kind of four different parts in my talk. I'll talk first about migrant connections, projects I've been working on at the German Historical Institute for the past eight years. And then a second spin-off project called Let the Past Speak. And then third, I'll talk about a kind of personal monograph research that I've been working on and how I've been using AI for components of that. And then finally, I'll kind of talk about the kind of field in general and what I

[2:14] kind of think of as a transition point we're kind of going through from, as you see here, from side glancing to vibe coding. And I did not realize when I was putting this presentation together that Lara Putnam was going to be here, but I certainly had her very important article from 10 years ago in mind when I was putting together my remarks. So for the past eight years, I've been the manager of a project at the German Historical Institute called Migrant Connections. It's an initiative to find, collect, and share correspondence sent and received by German immigrants to the United States over the long 19th century. And we initially launched the public-facing project in early 2019 using the Omeka Classic platform, which was developed at the Rosenzweig Center at George Mason, and incorporating into a crowdsourcing interface called Scripto to enable volunteer researchers to transcribe letters from 19th century German

[3:16] handwriting into machine-readable text. And we couldn't have anticipated it when we began, but now, in retrospect, it's not necessarily unsurprising that we experienced a big boom in interest in the project during the pandemic, as did other crowdsourcing platforms like the Library of Congress and Smithsonian. And overall, our volunteers up to date have transcribed some 1,600 letters. And so here's just an example of one of those letters written by a woman in North Rhine, Westphalia, which is kind of northwestern Germany, to her brother who had immigrated to the United States in the 1850s. And again, letters are an important source base for the history of the German diaspora. There are approximately 350 million letters exchanged between the two countries. And we've been working in collaboration with a project that was launched initially in the 1970s in Germany to collect basically letters sent from the United

[4:16] States to Germany. So, again, it's kind of like we're kind of been rebuilding the other half, the correspondence that flowed in the opposite direction. And kind of the long-term aspiration is to kind of meld the two databases together, their letters and ours, into something which can be useful for researching both migration history and the history of the everyday. And again, there's a lot of, you know, kind of really interesting, vivid material that is found in this. You know, people are usually writing to relatives who they may not necessarily have known that they would see them again. This is a woman who had a very difficult marriage, writing to her father, telling him, you know, the devil shall not trick you into thinking I had a good life. And this is actually a collection that we received, the digitized letters we received from a descendant of this family. And that's basically the woman on the lower right is Dorothea Schumacher, and that's her husband, Peter, with one of their children.

[5:20] So again, one of the ways that we kind of gathered this was working with libraries and archives to digitize their collections. If any of you are archivists, you probably know that there is a big archival debt, a lot of processing, a lot of things which go unprocessed. Again, I think we came along at an interesting moment when a lot of archives were kind of reconceptualized their mission and sort of were really interested in collaborating with us and you know basically we're digitizing their sources putting them online making them transcribable making them findable and again you know libraries have and archives have lots of collections which are not necessarily you know accessible and you know non-english language collections here in the U.S. you know tend to kind of go to the bottom of the pile so I think these institutions were really happy to collaborate with us when we said, we're interested in your foreign language collections. And so this is one event that we

[6:21] did in St. Louis in November 2019, which was really great. We kind of did an antiques roadshow kind of thing where we'd have people come in and bring letters and we digitized them. The woman you see sitting on the left is someone who transcribes German handwriting for a living. And so she was able to kind of tell people a little bit about what was actually in these letters that they had inherited that they had not necessarily ever read. And as you can see one of the things we did in collaboration with the German Society of Pennsylvania was you have these online transcription days where different volunteers would kind of help with transcribing letters that we had collected. And again the design of the project was influenced by the kinds of technology available at the time. So nowadays there might be arguments for introducing an automated handwriting text recognition component into the project. If we were, you know, starting from scratch, perhaps combining something like an AI -generated transcription with human quality control.

[7:24] And, you know, a couple of years ago, we kind of reached out to a couple of our volunteers and, you know, kind of to ask them what they thought about this kind of thing. And basically, their feedback was that what they really enjoy is kind of the puzzle of figuring out these letters for themselves. you know, having AI in some form kind of come along and supply that first was not necessarily appealing to them. But if you're interested in kind of thinking about, you know, in the intersection of historical research and crowdsourcing and AI transcription, I would recommend that you check out fromthepage.com, which is a company that is doing a lot on this front. They partner with a lot of museums and archives throughout the U.S. to kind of put their put their archives online and they have a platform for transcription and they're thinking a lot about how to kind of what a pipeline might look like if you're kind of starting with AI and then trying to get people you know moving towards publishable results. So more recently the GHI was fortunate to receive a grant from the Volkswagen

[8:27] Stiftung in Germany for a spinoff project called Let the People of the Past Speak. And I've been working on that in collaboration with Hans Boas, who's a professor of Germanic languages at the University of Texas at Austin. And what we've been doing is trying to reconstruct the sound of 19th century German. And we kind of have this thing we've been calling the dialect engine, which is kind of an AI-powered tool which leverages the original correspondence, dialect recordings from the 1950s, and other sources. to reconstruct the sounds of both 19th century German and then Texas German, which was kind of the dialect that emerged in Texas over the course of the 19th century as people from different regions of Germany, you know, migrated to the same areas and then each kind of started speaking their own regional dialect and then kind of variations and kind of synthesis emerged from that process. And so again, the kind of the longer term question is, you know, how did Texas

[9:30] German come to be and what are the different stages of its development look like? Team myself, Thomas Schmitz, computational linguist, and Hans Boas and his team at the Texas German Dialect Project, which has been they've been researching the Texas German for the past 20 years or so. And again, a longer term horizon kind of past the time of the project, which is coming to an end in January, is reconstructing spoken Texas German. And so kind of what we're starting with is trying to reconstruct one example dialect. And we focused on Hesse, northern Hesse near Kassel. And the different kinds of sources we've been using for this are handwritten letters from our collections, recordings. they're called the Zwirner recordings. That was the name of the scholar who kind of put them together, who interviewed different people to kind of capture what dialect sounded like in different parts of the country. The Wenker data is a set of

[10:32] questionnaires that were created in the 1870s when a linguist basically sent, he basically came up with a list of 40 sentences and sent that to hundreds of school teachers around the German empire and said basically, please tell me how your local language, your local dialect. And all those questionnaires came back in. And so you kind of see like all the different, you know, regional variations in spelling and syntax and so forth. And then also a dictionary of a specific dialect from the early 1900s. So again, we're doing this with a variety of AI tools, including natural language processing, machine learning, And kind of at the end is kind of creating this model which will be based on so TTS is text-to-speech and RVC is retrieval based voice conversion. And so again we're kind of starting with these letters.

[11:35] Our linguists kind of go through and kind of diagram what the different aspects of the dialect sounded like. and again this collection of about 5,800 different recordings of I'm going to skip that particular clip and so here's an example of a letter from the 1850s a man writing from a town near Giessen to his brother-in-law who had immigrated to the United States basically telling his brother-in-law I've got so much work to do and the difficulties that he's encountering with that Essentially, what the kind of technical process is basically kind of taking these recordings, which we already have, and kind of deconstructing them into kind of different tokens and then kind of figuring out how to recombine those in a text -to-speech format in order to kind of, again, reconstruct what German sounded like in the 19th century.

[12:37] and so kind of like the first round of this was kind of using a kind of off-the-shelf voice like you know kind of like a Siri type thing there's a company called ElevenLabs which kind of generated a German version of this so so up here in the upper left you'll see the first part of that letter and this is how a kind of off-the-shelf model. erfahren. So far, this is an

[13:42] example of a recording of someone from the 1950s basically saying, when I started school, the railroad hadn't yet been built in our town. So this was kind of like a source dialect, so to speak. So what happens here in the lower right-hand corner is where we kind of use retrieval-based voice conversion to kind of converge the sound of this person in the lower right speaking, you know, in their particular dialect with the kind of phonetic spelling. So maybe I'll just play this. This is the office shelf. And then again, this is the kind of combination. Lieber Schwacher, was mit eigentlich beweit dessen Briefs schriebe sollst du jetzt erfahren?

[14:44] So again, the idea is to kind of model the dialect mixture. So again, starting with this one particular dialect from Hesse and then combining it with kind of figuring out a pipeline for doing this with one dialect. to other dialects and then kind of melding those into something which kind of approximates, you know, what was like the first generation of Texas German, which emerges in the 1900s. What did that sound like? And there are recordings from the 1940s of people speaking Texas German. So that kind of gives you, you know, again, something else to kind of approximate with. So again, so what I've talked about so far is, you know, kind of one project which involves, you know, lots of volunteers, citizen scientists collaborating with libraries and archives. This project right now, Let the Past Speak, it's me collaborating with a linguistics professor, with a computational linguist. And then I'm going to move on to something completely different, which is

[15:46] building Baltimore Miniature. on and kind of talk about kind of my own personal research project. You know, again, this is kind of what I've been working on parallel with this other work that I've been doing. So I've been writing a monograph about the financial practices of immigrants to the United States between the Civil War and the Great Depression. And it's conceptualized as a national history, but I use Baltimore as a case study. And so for some of that work, I've done some mapping in what we might now call the old fashioned way. So for example, I have a chapter about the 1920s, where I look at how banks led by, you know, the kind of white Anglo American elite started reaching out to ethnic immigrant families as customers. And so you know, this is an ad, for example, from the Polish newspaper in Baltimore, another one from the Czech language newspaper, the Telegraph. And so you can't quite see it here, but this is like a group of banks down here

[16:51] kind of with this become Americanized ad. And then another from the Providence Savings Bank, which was one of the big savings banks in Baltimore at the time. And so I used data from IPUMS, which is a database of basically census samples, which provides demographic data to kind of map where different immigrant communities were living. And then also went through city directories and mapped where particular bank branches were opening. So if you know the geography of Baltimore, this is basically downtown Baltimore. you know kind of the main central branches central offices were and then these dots out in on the periphery kind of represent all the different branches that were created and then the the colored dots basically represent different communities like Russian, Russian-born, Italian, Polish,

[17:52] Czech, and so forth. And again this is basically done with like QGIS you know, kind of taking data out of IPUMS and formatting that in QGIS. This is basically, you know, kind of all by hand, so to speak, or by keyboard. Yeah, this was the 1920s. And so one of the things that I found was interesting. So this is a map of basically where the African-American community was in Baltimore. And so part of the analysis that I'm doing here is saying, okay, these banks were basically largely avoiding the black neighborhoods of Baltimore. But then I have kind of a different research question, which I'm interested in considering for the start of my book, again, looking at like the 1870s and 1880s, and looking at basically savings banks. So scholars have done a lot with savings banks. There's like Tyler Anbinder most

[18:52] in New York City and that was kind of like a big data project involving you know looking at lots of different account records. I didn't want to kind of redo that. But one of the interesting things I found at the records of the Savings Bank of Baltimore, which was another really big bank, was the records of basically like different societies and organizations. They were kind of their records were kind of kept separately from from, you know, kind of individuals. And so, for example, here you can see. So this is like a list of like the different institutions. You're not really able to see the highlighting. But, for example, there's like the Teutonia Bund is one example of a German Ladies Society. And then over here you can see this is the Germania Taylor Association. and so one of the things I became interested in was thinking about okay well when you have these kind of ethnic organizations that are getting

[19:52] together you know are these kind of people that are kind of assisting with each other in their neighborhoods or are they you know is it like people from like different parts of the city who are kind of coming together and saying oh okay you know we kind of recognize ourselves as like a community and let's get together and you know start these different associations. You know, often it is kind of brotherhoods, fraternal organizations, women's sewing circles, and so forth. They're kind of seen as, you know, possible competition or prototype for a different vision of, you know, welfare and mutual aid and so forth. And so I was kind of interested in, again, this question of like, is this sort of purely ethnic association which kind of transcends the whole city, or is that something where it's like people in different neighborhoods kind of banding together? But to do that, you kind of need to know, you know, where were people living? And so this is, those of you who are American history scholars know that, you know,

[20:57] the 1890 census disappeared, one of the great tragedies of historical research. But, you know, for the early period, there are the 1870 census and the 1880 And other researchers have basically gone through the 1880 census and kind of mapped completely where people lived in the largest cities in the United States, including Baltimore, which hasn't been done for Baltimore. And so I sort of started off thinking about, okay, how can I do that? Can I kind of look at like, you know, where are these societies? Where are the people who are associated with these societies living in 1870? And then where are the people associated with these societies living in 1880 and kind of compare and contrast. And so again, if so, one of the people who was mentioned in this text here is named Robert Brookhoff. This is him right here. So I searched for him in the

[22:01] directory and found his name here. Of course, this would be incredibly time-consuming to try and do this person by person by person by person. So this was kind of like one of my first runs of this was kind of going to Google Document AI and basically kind of giving it a model of how I wanted it to kind of go through the address directory, which has been digitized and put on archive.org, you can basically kind of define a data model and say, okay, basically like, and it's pretty consistent that they use commas to kind of structure, here's the person, here's where they lived, here's what the occupation was. And so you can kind of basically go through and kind of train a model to kind of recognize these things and, you know, sort of tweak it and, you know, basically get going on that process of like being able to actually say, okay,

[23:05] here's like a database of like, who are the people living in Baltimore in 1870? And what were the addresses? And again, this is again, you know, on ancestry.com, I found Robert Brookhoff. And then in IPUMS, you can kind of pull out the data and, you know, be able to look at, you know, like kind of individual people. And again, he married, did he have children? You know, what were the kind of the characteristics of these people who were involved in, you know, these kinds of associations that I'm interested in? But there's one issue, which is that, so I'm looking at 1870, I'm going to look at 1870 and 1880. But Baltimore changed its street numbering in 1887. But they also kind of created this guide in, like, the first directory that came out basically saying, okay, here's where the old house number was and here's where the new house number was. and so what I kind of wanted to do for this

[24:07] presentation was like, you know, I think there is kind of talk through like kind of like the process of kind of kind of doing this kind of research question and saying, okay, what is my primary data and like how am I going to try and take that and like go to the thing I actually want? So this was like, okay. Oh, and the other thing I should mention is that basically the that in the 1870 census data they don't include addresses they do in the 1880 census which is why that project was able to do that but basically the directory the city directory was kind of used to find people for the 1870 census so like that's how you know you can use the directory to kind of do the geocoding of these people But again, there's a street directory with street numbers. And for example, Aliceanna, I just love this street name, so I like bringing it up. But this is Aliceanna, which is kind of in the Fells Point neighborhood. And then there's this kind of list of

[25:10] old and new house numbers. Maybe not be able to see it well, but basically it has like the new house number first, then an asterisk, then the old house number, and then it also gives you the names of the cross streets. So that's basically saying, this is Caroline Street at one end, Caroline is the cross street at one end, Dallas is the cross street at the other end, and here's a list of all the house numbers that changed and how they changed. And so what happens is that I have a conversation with Claude about restructuring the street directory information and I kind of say okay you know here's how here's how that you know that list of like house number changes is defined and I want you to go through it and basically give me a block by block reconstruction that kind of says this was the street name these were the two cross streets this was the old name this is the old house number set of old house numbers this is a set of new house numbers. And basically, this is kind of what it generates.

[26:12] And I mean, of course, this is not perfect at all, but it's actually like really good. Like I think basically one of the things that I kind of set it up to do was to kind of give me a list of like, here are the things I couldn't figure out. Like, you know, because some of these numbering schemes were a little idiosyncratic. But like it's on the order of like three or four errors and it's about it's like 60 pages or so of like directory you know house number change information so you know it's it's kind of tedious but doable you know it's like something you can do in like a few hours and so what you kind of end up with just to kind of visualize it is like basically again this is in QGIS to kind of map it out is basically saying okay and sorry that he's all turned out the wrong way around. But basically, it's saying, okay, this is, it starts with like, here's what the old house numbers were, and here's what the new house numbers were. So like, the kind of long term plan is to kind of

[27:17] take that directory information that has addresses extracted, and then be able to say, okay, well, I know that this person lived at, you know, 400 Aliceanna Street, but like 400 Aliceanna Street is now 900 Aliceanna Street. So put that on the map for me. And then go on from there with the rest of my analysis. And so again, this is all kind of based on like, you know, kind of vibe coding and kind of talking back and forth with Claude, with Document AI for the other stuff. And some of it's been working with Copilot. And so, again, I think one of the things we're kind of thinking about is, you know, Lara Putnam wrote this article a decade ago about, you know, the transnational and text searchable. if like someone can sit in their computer here in Maryland and read, you know,

[28:18] newspapers from Costa Rica, newspapers from London, newspapers from Panama simultaneously. And, you know, it kind of can give you this sort of sense of, you know, kind of comprehending all this information all at once. But like, you know, you're also not seeing what isn't there. Right. And so I think there's a kind of like the dilemma that we're facing a lot of us as historians, which is that we know that like vibe coding is like a thing where we can kind of answer these new kinds of questions and like kind of figure things out. And, like, it can be something that happens really fast and, well, not fast, but, like, you know, you can kind of do something like this, like, at such a large scale so much, much quicker than, you know, you could have, like, if I was just trying to do that with, like, the way I did that project about banks where, you know, I was only looking at, like, 40 or 50 bank branches and putting those on a map. so um so again i think this kind of changes the way that um historians are

[29:23] especially you know those of you who are kind of come here because they're sort of motivated and thinking about these things uh are able to do um able to do history and so i think one of the things uh which kind of came up in uh the the morning plenary session was like you know this question of like, okay, what are we going to do now about, you know, like, what's the next digital history lab going to look like or something like that? And I mean, I think that there's a kind of interesting question that's going on right now, which I kind of say is like, you know, maybe we are, I think what happened about 10 or 15 years ago was that there was very much this idea that like, let's kind of do these kind of big projects. Let's have these digital humanities labs. Let's kind of be, you know, the people who are doing big data and data science and it kind of will kind of adapt, adopt the terminology of, you know, the kind of big science funders and big engineering and big medicine and so forth. And that that way we'll kind of ensure a kind of institutional survival for

[30:27] ourselves in the context of, you know, how the university system is being turned upside down, not only in the United States, but also, you know, Great Britain and Germany and other places. And so I think we're sort of seeing that, you know, that idea of the lab as the thing that will save the humanities has not necessarily kind of come to pass. And that, you know, there are, you know, wonderful digital humanities centers that have, you know, seen changes in the funding model, funding opportunities. And so, again, I think that like what we may be thinking about, what we may be encountering is kind of a change or a shift happening where, you know, we may need to start thinking again as like rather than a lab, you know, a historian who's part of a lab with like lots of other collaborators and so on, thinking about like the historian who's at their workbench and like, you know, has all these kinds of tools to do the kinds of projects that like they couldn't have really done on their own five or ten years ago. So, and so I think that it's just sort of, I wanted to kind of raise that kind of

[31:28] question of like, you know, what kind of moment are we in and, you know, how AI is something that can kind of offer these opportunities that like, you know, that may lead us in a different direction than we thought, you know, digital humanities and computational humanities was going to go five or 10 years ago. But I think, you know, it's just something to think about as like a kind of alternate path that we might be on. And then just a couple of kind of useful things I'll just mention. So Cameron Blevins is a really great digital humanities scholar, wrote this post on GitHub a few months ago, which kind of where he talks about like a particular data set that he had kind of looked at briefly during his dissertation research, I guess, and kind of set it aside and then kind of went back to it and, you know, kind of spent an afternoon vibe coding it and kind of saying, you know what, it wasn't really worth doing anything with this, but at least now I can kind of figure that out much faster than I would have otherwise.

[32:28] And that's kind of like a good, perhaps a good way of thinking about, kind of get the creative thoughts flowing about how you can kind of apply that to the projects that you're dealing with. And then also I'll just mention just in the sense of like something that's out there for those of you who are specifically interested in immigration history, to history of migration. The ship passenger records, which are basically kind of taken from the manifests of ships arriving in the United States, are basically in, have been posted online by NARA in a very endearingly clunky interface, but you can basically download all of those. It's about 11 million records, I think. And again, it's the kind of thing where you can kind of say, hey, maybe I can kind of experiment with this. Like, you know, what can I find out about, you know, people who immigrated in the 1870s? What can I find out about, like, you know, people who immigrated with, you know, younger children versus older children or something like that?

[33:30] And again, this is the kind of thing where, like, you can kind of use Python to kind of, because the way the records are right now, they're, like, divided by country or empire, more or less of origin. but you can kind of combine them into a massive data set and kind of reconfigure them and just sort of think about what's something that can kind of help me think about how I might want to work on big data, but at my own personal workbench. So I just commend that to you. I'll also mention my colleague, Jana Denskeck, who's at the German Historical Institute, just published a book on basically the German-American newspapers and does really amazing stuff with digital humanities. And she's looking at German-American newspapers in the 19th century, basically, and looking at kind of like the syndication of articles and poems and other things like that. And kind of, again, started with a data set from the Chronicling America project and then did a lot of computational work to kind of

[34:35] trace how different texts move around and circulate and, you know, how different ideas about gender kind of circulated along with these texts. so I think I will stop there and I hope this has been helpful and I look forward to your questions or conversation let me just check if I can, is anyone in the discord uh huh how do you handle a routine that you're fabricating the path or that might ask what it means ethically to take human's voice and have a machine reinterpret them obviously the research matters but i suspect some folks offer that type of yeah i mean i think what uh yeah i think we try and talk very consciously about the project that's something that's like reconstructing and

[35:39] it's not like recreating that you know yeah this is like an artificial reconstruction but we think it's kind of an authentic it kind of gets you towards what something someone sounded like back then and again it's kind of like a one way of kind of trying to think about how language changes over the past by being able to hear it as opposed to you know simply reading it on a page yeah and and I think like basically in the final version project what we're intending to do is actually have someone who's from like that region be like kind of like the source source voice so to speak uh and then kind of have them like like give them like a a few i think i think it's like basically like a 30 minute kind of training set or something like that and then that would kind of be what we use to kind of generate the kind of uh voice that can basically read whatever text you give it in the voice of that dialect. Yeah. Yeah.

[36:40] I mean, music theater recordings is another source. No, that's an interesting idea. Oh, uh-huh. Yeah. As the immigrants that, you know, you may get more more dialectic humor. And, you know, there may be ways to kind of map where those performers were coming from as well. Yeah. Yeah. Thanks for that suggestion. Yeah. Yes. To add to your second project that you talked about. I was an undergraduate a million years ago working for the Parfives at the Birmingham Public Library. They recovered from storage under the seat at the stadium. A huge number of German language newspapers from 1872 to World War I. I was just looking at their website.

[37:42] I don't see any indication of it. Most people don't realize that Birmingham, Alabama was essentially a planned community that the steel industry built. And there were a huge number of jobs. And so it was a huge destination for immigration. And there was a tremendous German community there. And I don't know what they've done with them, but if you can get access to them, just, you know, to a comparable place. Yeah. What would be something that's in Birmingham? Yeah, thanks for pointing that out. Which was the original Magic City. Not Miami for it. Yeah. There was a project that sourced out of the University of Ottawa, some Quebec Sun a decade ago, where they took political speeches, recordings, so it could be music. Yeah, they had spoken word too. spoken word political speeches so you might have dialect in your moves and

[38:45] it was an experimental web page project where as you were navigating on a QGIS or equivalent now the sounds of those speeches when you were in that location you would reverberate they would be co-mingled with speeches from other locations in the city. And as you moved, they were trying to reintroduce sound in HTML but kind of create a feel for dialects and people and location in the city. So as you moved in between two neighborhoods, you would have two sound streams coming at you and overlap, et cetera. I forget how they use it. more bringing sound in to create, not to recreate individuals cybernetically. Yeah. Which is what Discord person seemed to be concerned with, but to

[39:50] add some authenticity to the spoken word. Yeah. My question, you kind of introduced us to several different kind of approaches that you've taken different methods. I'm curious about what maybe in observing these approaches surprised you or like going through each of these projects. Is there a move to, you know, your most recent project when you're noticing something that's been surprisingly difficult or surprisingly easy or like that you didn't expect to kind of learn? Or I'd love to hear kind of your reflections about the process and what that kind of felt like. I was surprised at how easy it was to basically do this house numbering project. Like, you know, basically kind of saying, oh, I have this concept of how I think the data should be organized and, like, can you actually do it?

[40:52] And, like, actually, like, when, like, the first run of data was, like, really accurate, I was like, oh, okay, this is, like, not, you know, this is not, like, kind of a pie in the sky thing. And I mean, I think part of it sometimes, I think, is just being able to say, I want you to do this data. I want you to process this data. And this is exactly how I want the output to look. And it kind of gives it something to kind of start off with. So I'm not just kind of saying, oh, please make a database for me. I'm kind of saying, oh, I want to have these fields of data in your output. And then it kind of figures out and kind of telling it, these street names are in parentheses. There's like look for the asterisks and so on. So it kind of like just being able to kind of get it to kind of do that was kind of like, oh, wow, this is like, you know, not as there weren't as many steps iterations.

[41:52] Yeah. Yeah, yeah. yeah yeah and i mean i i think um that was another part which was kind of actually that was the other part was that like that it did not hallucinate numbers like that was like kind of the thing that i really worried about of course is like if it just kind of is adding numbers and just like oh i think this should be this i mean like what what was What it did do is like it kind of, I don't think I included the column in the screenshot, but there's basically like another column over here where it kind of gives some notes that kind of say, oh, in this case, you know, I think there was a missing asterisk, but it doesn't affect your full data. So, you know, but just so you know. So it's like I can kind of look at it and say, oh, okay, here are the things that it found. the things that it found that are idiosyncrasies and they match what's there it's

[42:53] kind of like okay it's not kind of going off on its own and making things up yeah um basically suggested a column for irregularities right so it suggested if i can't if this doesn't at what you told me to do then I would make another. Yeah and the other thing is that it can like basically it kind of lists the like so what I was so like I said it's about like a 60 or 70 pages of address information but I was having a process at like three pages at a time and then each time it would kind of spit out okay these are the things that I was unsure about and I kind of numbered those for me and you know I kind of after a couple of tries it was kind of like okay you can like say you know number 25 yes do what I say number 26 you know yes go with my correction number 26 ignore what I how

[43:55] I suggested you correct it number 27 if it's something entirely different you can type that in and so forth so again I kind of yeah if it's uh if it's cooperative I don't know I mean these are like anthropomorphic ways we think about it. But you can have these things built into it to say, here's how I want you to deal with fuzzy data, and here's how I want you to not just go off on your own, but let me know that's a decision point for me. Yeah. the rectory itself had like an overview of the street name change street numbering changes and that was part of your input to be able to say here's the algorithm yeah and were there any other sources like the media people you know had to be told that

[44:57] their number was changing no i mean there was a city plan that said oh we're going to do this uh-huh and here's why is that some of the background that would inform that um not really i mean so i mean the thing that's kind of really fascinating about like the 19th century uh street directories and i think this is found in other uh cities too is that like they would basically give you a list of like they they would tell you the street name. Like in the back, there would be like a list of street names, and then it would tell you like block by block what the house numbers were. So it was like they were already kind of capturing that information year by year. And then it was like, but when, you know, all the numbers change at once, it's kind of like, okay, well, we don't have to like start with kind of building a database of like how all the house numbers in Baltimore, we can kind of start with the information we already have, and then kind of apply the rules that have been you know implemented by the city saying you know

[46:00] oh I think there's someone named Reuben Rose-Redwood who I think it's kind of like the main scholar of the history of house numbering although I think he's worked on New York City mostly but I think this is something that was like you know sort of early progressive era rationalization of like you know like basically they kind of switched to a system of like each block having like a being numbered in the hundreds instead of it being like you know 1 to 13 on one block and then 15 to 29 on the next block it was like 1 to 100 101 to 200 and so on like that's like the Philadelphia system which had been there like since the 18th century like a lot of a lot of cities kind of basically established that or adopted that yeah yeah yeah Yeah. The city names in the town and towns get changed according to which side of the political spectrum won the funds. So you get these little criminal publications that list the house numbers, but also the names.

[47:03] So it's like the despotic. Right. From the street names. Yeah, and like in the final version, it's all I have to like kind of incorporate data, which has basically been gathered by like real estate people of like street name changes. Like for example, there was a German street, which was changed, I think, to like Redwood Street during World War I. So it's like kind of like, you know, sort of have to add those into like the actual final product. Yeah. Yeah. What's that? Yeah, so basically the idea is that I want to kind of look at, you know, where are these people living in the 1870s?

[48:07] I think I probably will do something that's kind of like taking like a sample of people who like individuals who opened accounts and kind of saying, okay, where were they in 1870 and where were they in 1880? And like the IPUMS database, which I talked about before, like they've done a big project, like basically to find, to link people across censuses. So I don't have to do that myself. Thank goodness. Yeah. No. Oh, I see. That's, yeah. No, that's a really good point. And actually, the thing about savings banks is that they were kind of conceptualized as a place for like, you know, widows and orphans

[49:08] to be able to keep their money and to have access to it. So you do have a lot of women opening accounts sometimes, you know, like women opening accounts for their children, for example, or, you know, women who are, you know, single in the workforce opening accounts for themselves. Yeah, and I think that probably that will be an issue to kind of figure out like, you know, if there's a woman who opens an account when she's unmarried in 1870 and then 10 years later gets married. Yeah, how do I how do I find her? Yeah. Yeah. Mm-hmm. Yeah. Yeah. Yeah, yeah. I think, yeah. So I think like in a given year, you'd have, you had about like, 3000 people or so opening bank accounts. So yeah, I'm still I'm still figuring out. But again, that was another reason why I wanted to kind of focus on looking at

[50:09] these societies, some of which were, you know, organized by women themselves, you know, to kind of say, okay, yeah, like, if I have, like, you know, I think it's something more like, you know, 50 or 60 societies or something that open accounts. have like 10 of those were run by women and they have like you know the accounts the societies usually had like three or four people who were kind of like the board of directors who could open the account so it's like yeah okay if i can find 40 women though that that number i think i can track yeah what's left the german society records that individuals have found and uh uh publishing is that that could be a really interesting source yeah yeah um now I'm not like kind of solely looking at uh German immigrants for this uh project but yeah it actually is the case that like at the Maryland archives they do have like a record

[51:09] book of the German Society of Maryland um a kind of uh listing like who you know who like applied for relief and and things like that um but like the quality of the scans was really terrible but it was it was a few years since i was there so like maybe nowadays i could like get a better better version of that or something like that yeah yeah your discussion of um you know this transition from like big digital humanities projects to more of this workbench model, it has me thinking about issues of credibility, recognition, what the kind of state of the field for historians and archives and librarians in this AI model. And so I'm wondering how you

[52:10] kind of tie this morning's conversation about the state of the field to that kind of big project, big team approach toward the individual workbench, you know, archivist, historian, librarian, and where we can insert individual citation, individual recognition into that? That is a really good question. Yeah. I mean, I... Right, right, right. Yeah. Credibility and therefore recognition pay that we all feel we deserve. Yeah. And, you know, so then we're recognized as legitimate experts. Uh-huh. Yeah, I think one thing that people have talked about is, like, kind of having data sets be more recognized as like a research product that's like kind of important and like have the idea that

[53:15] like, you know, like if I end up with this kind of like geo database of Baltimore in 1870, it's like not something that I need to keep to myself. I can like kind of pass it on other people and like kind of show them, you know, all the different models and code that I use to kind of generate it. And then like it's like, okay, the next person who wants to answer a whole different set of questions about Baltimore in 1870. 70s can use that and you know kind of cite me for that yes i think well uh-huh big science model yeah in the kind of like you know science fields the pi gets cited but all the little grad students don't necessarily all get cited and so it's like that's where i think using that that model and also you know kind of recognizing where the errors get in I mean I'm dealing with you know archival data that was you know created from an early 1990s digitization project

[54:17] and metadata sucks it's all sorts of errors and it's like how did those errors get in it what were the biases and you know the blinders that then caused those errors how much of it was just like a bad archivist or how much of it was whatever, you know, kind of computational tool that they were using in terms of early, you know, just the structure of an archival management system and how it's translated from one to the other. You know, where do those errors get in? And that kind of leads to this question of what is the validity of that metadata that then we are relying on now that we don't, you know, recognize where those errors come in. Yeah. And be able to, you know, what's the authenticity and the credibility of the people who are creating metadata? Yeah. The people and the tools. Mm-hmm. We got to sign them all. Undergraduate student workers. Yeah, yeah, yeah. Yeah. Yeah. Same thing happens with the recon projects with cataloging when people went from car to digital.

[55:20] Mm-hmm. And then it's leeway, right? So some places, like Ivy League schools, the library basically said, any record within 10 years of the pub date on the card, you can feel free to use that record. So knowing sort of the process of generation of that stuff, there is no, I mean, and so recon projects to me always like going from a card to just input the card. That never made sense. You need to look at the thing. Right. Yeah. And so that, and of course, there's no funding for that. There's very little funding for remediation. There was very little funding back then, but now, of course, we're finding we have to remediate. Yeah, I mean, where I think archives are, because of the whole reparative description movement, that their remediation of just bad data is coming along with that. So, I was hired to do reparative description, but half of what I'm doing is cleaning up just ugly data, names that got split into misspellings and things like

[56:20] that. worse could have been student workers could have been character limitations in the you know could be pulling in you know copying and pasting something that's already problematic i mean i just did a stand-up two years as a metadata librarian although i'm a cataloger and a medievalist and a cataloger of my trade um and so i'm now a distinctive the distinct collections cataloger at Arkansas, but when you go to the source for your records, I am always having to do all of those things. And reparative, enhance, update, reformat a little bit. I just want to get... The cataloger was, is as important. Well, given the strange fact that apparently the cataloging staff where I used to work So the state of the catalogers, the metadata librarians also makes a difference.

[57:24] Yeah, cataloging metadata, it's a power tool. We don't use power tools when we're altered. Can I get one last question? Yeah, I just really appreciate your thinking in these terms and thinking about like what is, you know, side glancing took us in one direction where does by putting takeoff? And it's interesting to think about how source digitization, you know, 10 years ago, 10, 15 years ago, especially made visible new geographies. Like suddenly we were searching, you could search without knowing where you were looking in geographic terms. but and so many of the super interesting examples that you're pointing us to are playing a foreground the fact that you know if i've got to have the capacity to work with you know lowering the barrier for the kinds of transformations you're describing especially helps us like structure unstructured data or move from structured management data to structured digitized data yeah um like that that

[58:28] barrier goes from like you need a whole fleet of undergraduates to it can be in this workbench where they're describing. And so what that makes especially, it lowers the cost of changes of scale more rather than lowering the cost of side glancing across geographies, right? And so then that's really interesting to think about how is history well positioned to take advantage of that, like increasingly structured data? Well, one thing you could do is structured data is quantitative analysis, but most historians are not trained in that way. Some archivists maybe, like is this a case where archivists data may end up being more source of new analyses or other disciplines. Yeah. Pushing in these directions seems to be really fruitful for us to be thinking about, as we understand the transformations that are not going to take place as a cutting edge, but rather as what's easy shifts. Yeah. Yeah. Yeah. No, I think, I think, yeah, it's very much about the change in the barrier to entry. And like, just again, the idea that like you can, you can

[59:29] sort of imagine doing things with of seen and knew there was something there was some information in there but couldn't quite figure out how to get it out of there and now it's kind of like the barrier just to kind of saying like okay I just want to like put this you know I want to describe to a model how I think this data should be structured and see what it comes up with and then all of a sudden you can kind of say oh now I can answer this kind of question which I thought was probably inherent in this data, but was not actually what the people who collected this data were looking for. Yeah. Yeah, if anyone's interested, like, you know, there's this big project called the US Immigration Commission in the early 1900s. Katherine Benton-Cohen's written about it. But there was one of those things where they have like 40 volumes of, you know, different, looking at different communities all around the United States. And some of those volumes are just nothing but

[1:00:30] tables. And I think no one has actually looked at those tables since they were printed. And it's like, okay, what could someone do if they kind of went in there and said, okay, I want to like take that data and like actually be able to compare things that the people who are putting this data in the book did not actually want to be compared. And if they realized that someone 90 years from now or 100 years from now could compare they would have said, oh, don't publish it at all. You know, so. I don't know, too. It's the when the FDR administration took office where it was recently established a central statistical bureau. Basically, and there's a volume describing all of the new agencies in the new deal. And when it describes creating the Central Statistical Bureau, it talks about all of the previous sort of departments and organizations that were

[1:01:32] collecting statistics that in ways that were basically not valid and rendered everything not comparable. and so that this the central statistical bureau was established in 30 33 34 to regularize and to be to consult on all of those departments and entities collecting statistics makes me what you just said makes me think they probably looked at that project oh my god uh well thank thank you everyone and I think it's uh we're a little bit over but uh yeah I really appreciate the conversation Thank you.