 And we move on with the next presentation, which will be done by Julian Hemchen from Freie Universität Berlin. Hello, good afternoon. My name is Julian and due to the format Lightning Talk, let's jump right into the topic for today, which is the creation of a general model for German handwriting from the late Middle Ages, namely 15th, early 16th century. And that's important. What I'm presenting right now is by no means a finished project, but more or less a request for aid. This is because I, as well as many other medievalists, I assume, was facing some hoops to jump through when starting off with working with transcribos. For my own project, which is called Mapping Medieval Vienna, I need to transcribe more than 3,000 pages of Vienna's real estate records from the 15th century. And without transcribos, this would just not be possible within a timely manner. Yeah, but to really take advantage of the HDR tool, I needed to train a model from a manuscript from scratch, and that was rather time consuming. This was because at that time when I started, which was in 2018-19, there was just no fitting base model to draw from. So most models back then were covering mostly more recent German from 16th century onwards. Yeah, and in the end, my own model turned out to be functioning quite well, but only for my very specific sources. Then last year, Angela Wang and Gerhard Schwedler from the Universities of Kiel and Lübeck hosted a summer school centered around transcribos. We heard of that before this day from Vivian Popken, thank you. Yeah, and when discussing our ways of using transcribos, we discovered that many of us ran into the same kind of problem. So, and that then kindled a spark in us. What if we take all the training data gathered in our individual projects and just mesh them together? In this way, we can create a more generic model that can then be used as a base model for all future models to be trained. This would have all other medievalists to get started and make the access to late medieval German handwriting for everyone more easy and accessible. Yeah, to put this idea into practice, the group from the summer school is more or less meeting monthly to discuss our individual progress and contribute more ground truth to our group. Yeah, so far we have created a collection where we all link our material to and based from the linked transcriptions, we created the first model prototype from a training set consisting so far of 70,000 words. That's not very much. Yeah, but yeah, the method used was HDR plus and the validation set or the CR on the validation set is 3.8% pretty good so far. So yeah, here you can see what it gets right and what it gets wrong. Maybe the CR is for real a little bit higher because the validation set only consists of material which we have taken from our sources. So that's the only page on that validation set which is like not linked to our projects and it's still functioning pretty okay-ish I would say. It could be improved but nevertheless I don't want to play it down. Yeah, so the next step would be trying it with more new material and testing the Pileier engine. But in fact we don't expect the error rate to drop significantly because just medieval documents have some features which will eventually cause problems. For instance, different German dialects, low German, higher German or different kinds of scripts and fonts even in the same document. Like for example, all these are taken just from my Viennese records and then a big point the highly varying kinds of abbreviations. Yeah, to solve some of these challenges it would be helpful to agree on some transcription guidelines and maybe hit the lowest common denominator there. So far we have not done this because more or less we are all like, yeah, we must stick with our individual guidelines for our individual projects. But and that's the point we hope to even out this challenge and making the ground truth as diverse as possible with the following idea. Yeah, it's pretty easy. To improve the model we need more data. To get more data we need more transcriptions and transcribers. And the better the model gets the easier it is to start transcribing. So that's the big point. If you are working with late medieval German material or have plans to do so please get in touch with us, namely me and or Angela Wang. Yeah, we would be very happy if you would join the group and discussing all transcriptions or transcription related topics. Yeah, and that's also important and the transcriptions itself are not made public so all possible rights will remain with you. Yeah, and last but not least if you want to try out what we already achieved our first prototype is online since last week. Thank you Miriam. She's not here. She made it possible. Yeah, you can find it with a link or just on the transcripts website. And so yeah, let's join forces. That was the title. Thank you. Perfect.