Transkribus Field Models Webinar (English)

Okay, let's start. Welcome everyone to today's fields models webinar here at Transgrievous. We are happy that many of you have joined today's webinar and we'll have a look in the next about 60 minutes what field models are and give you a good introduction into these models in transcribus. As you know transcribus, you probably have worked with similar documents. Today when we talk about field models, we will go a little bit beyond what you are uh probably capable of doing with transcribers already. So doing text recognition. Today we will have a closer look to field models and also answer some in-depth questions about using field models. Just a very brief overview. First we will start with a quick introduction. Then we will mainly talk about preparing training data. So that's a very important and fundamental step. When you work with field models, you need to understand how you train your model based on the data that you are able to prepare. Then we will have a look at how the actual training works, how you start the training and what you can keep in mind in terms of settings when doing a field model training. And then eventually how you can use your trained field models and at the end we will have some time for questions. As always, if you have questions, just type them in in the chat during the webinar. So we are also happy to take them if they fit during uh the things we are discussing. uh but there's also time at the end to answer your questions if we were not able to answer them during the webinar. Who's here with me today? First we have Helena from our com team who she's leading our marketing and coms uh team and yeah does a great job in preparing all of these things for the webinars and our uh presence online. Then we have Sara here from our customer success team who's yeah managing many of our very successful customers when it comes to executing larger projects and really doing the fancy stuff with Transcribus and yours truly. So I am part of the board of directors here at Transcribus and also leading the product development with Transcribus. As you've seen uh Transcribus can easily read such documents. this very example, this very clean handwriting as well. So reading this text or recognizing this text is not a problem for transcurous at all. But if you have a closer look and we have a little bit more complex layout in this example, you probably wonder how can I untangle this? Because if you have a look at the first uh line and then the second one, you already see okay that's a heading and then we probably have a page number and they are in line number one and two and then there's the third one and then if we have a look at uh six and seven there already things getting complicated because these uh uh reading order does not uh yeah provide a lot of value for you as a user as you would like to have those uh sections of the text uh untangled. So, we know where the text is and what it says, but we don't know really where it is and how it uh is related to each other. Here is where field models can come in. Uh you can probably think of them like a cookie cutter. You can cut out different parts of your documents and assign labels to them. So you can automatically detect those fields as we call them with field models uh and assign regions and text to them as you see in this example. There's a heading left of that heading and paragraph section. There's also marginelia and yeah the main text is in the paragraph. So how do we make that work for you? Today we will have a look at how to prepare training data as said before how to train your model. How you evaluate the model that you've trained. So to understand am I training a good model? Can I use this model for my production phase now or not? And how to apply your model. So we hope that at the end of this session you will be able to do all of these things and can successfully work with field models in transcribus. With transcribus as you probably know we try to really uh deliver uh an AI ally to you that you can use for to simplify your time consuming and laborious work so you can focus on the fun part of working with historical documents. So how does that look like in practice? Now I will hand over to my colleague and we will start having a closer look to the actual work when you work with field models in transcribus. So take it away. >> Thank you F and we will jump right into the topic. Just make whole screen. There we go. All right. So as Flo already explained, field models help uh transcribus identify and categorize different layouts element layout elements in your in your documents in your material and make sure that the structure is preserved and also understood in a semantical manner. So up until now when we used layout recognition in transcreus uh we focused more on detecting the text regions and then the text baselines and then extracting the text from there. So we did get the transcript but we didn't really know uh if the text was part of just a block of text or marginelia or if it was a heading. And with field models, we can now train transcribers to automatically recognize and also tag these regions and therefore also to specify this additional information in the transcription. So with field models, we'll then have to do some additional steps compared to the text recognition. Um but the benefit is that you get this additional information about the layout of your document. So when I talk about additional steps, what do I mean? So um unlike with customtrained text models where you train a text recognition model and once you use it, you get the regions, the lines and the text all in one recognized. With field models, we need to split those steps up. So we need to make sure that really only the information is extracted together with the specific layout elements that we want. So what we recommend here is to look at your documents. Look at what kind of documents you have. What kind of structure is in these documents in the material? What does the layout look uh like? And do I need to use a field model or maybe do I just need the normal text recognition model? Of course, you can train a field model for different types of layout even if there are just some small variations and deviations that you want to work with. We have some examples here. So, uh when field models are really useful is for example for newspapers. Uh you can train a field model to really recognize those paragraph indentations to recognize when a new paragraph starts. Um you can also use them for legal forms, index card, court records. So really a large variety when field models are useful. We can have a look at some other examples. So here you see an image where it's really mostly about different text regions but again specially useful especially useful uh it is for newspapers um where you can train to recognize the specific paragraphs. You can see here this percentage that is next to the uh paragraph tags and this indicates the confidence level meaning the confidence uh in which transcribers recognizes those paragraphs but we will talk about the confidence level later uh in more detail. You can also of course recognize uh just different columns. So, a simple layout with two or three different columns. Um, and you can also recognize uh the content of forms. So, here it's really interesting and useful if you only need specific information. As you can see here, for example, maybe you don't need the text that indicates name. You just need the information of the name itself. And in that instance, you can really just train the model to categorize these layout elements and then in turn only extract the layout elements and text elements that you need for your work. But of course, it's not just limited to text. You can also train the model to recognize, for example, illustrations, as you can see here on the left side, um or even elements of sheet music. And we actually have a project in Austria where they are working with sheet music. So there are no limits and there are some very exciting projects um that are working with this technology. Now to preparing the training data. So now we have our documents, we have our let's say our forms um or our newspapers and we want to prepare the training data to train the custom model. How do I get from my basic layout to actually the training data, the ground truth data to train your model? Um, if you've used transcribers before and attended a previous webinar, you might know this workflow. So, generally speaking, we have different workflows where you can also use public models. For trainable layout models, we often have only a limited amount of public models such as also for uh field models. We do have uh a few. We have for example two public models that we trained. The baroness of blocks which works well as the name says for blocks and the marginalia monarch which works quite well for marginelia. We had a few questions in the registration form asking how to recognize marginelia. Of course, you can train your own model to recognize marginelia, but the marginalia monarch is already quite good. So that is something that we can also recommend trying out. But if the public models that we have, we have also a few published ones uh from users. If they're not sufficient and you have specific layout that you want to train your own model for, then of course you can use the next workflow that we'll show you. And for the training of this uh of this custom model, as usual, it's always very important to have specific examples that we can show the AI to help transcribers recognize and understand the layout in your training data. So you need to prepare these examples. And how that works preparing these examples is to first of course you upload your documents. But then you accurately draw and tag the text regions that you want to train your model on. Then you save these uh pages as ground truth and you start your first training. With field models, we recommend to have uh 50 pages of ground truth as a start to train your model. and I'll show you how you can prepare these examples in the web app. Let's have a look. So, while we're here and I have all these examples, maybe first uh I can show you an example of public models. I have one page here where we can see on the left side marginelia. This is an example where it's already recognized. But let me just go one step back store this version. So this is the empty page and how you can use the public models is to just click on process with AI and see this. This looks funny. Okay. So, do you know why it's so squished inside and not opening properly? >> We can see uh start recognition at the bottom. >> Okay, perfectly. So, it is important here when you're working with fields to switch from text to fields and then here you can select the public recognition model. So in this case I selected the marginelia uh of um no yes marginelia monarch but you can see here there are already uh quite a few models uh that are also public as well as some private ones that we've already trained but if I select now marginelia monarch I can see I've selected just one page and I start the recognition it will as you already have seen the preview before it will only recognize the margin marginelia in your document. So it will not recognize the blocks of text but really specifically just the marginelia. Now we are number one in Q. But I can just go back to our previous field recognition. You can see here in our version history that I used field recognition. And to just show you Oh, it's already finished. You can see here that low that maybe you can see here that really only the marginelia has been recognized and not the blocks of text. All right. Now if I want to train my model and I need to prepare the training data, what you do is you again open the editor. And in this case, you can as usually see the layout editor on the left and the text editor on the right. For preparing the training data, we really only need the layout editor. So I'll just enlarge in that a little bit. And then you want to start with drawing fields and regions. And you do that by opening add region. Then click on add text region. In this case, we want to add text regions. And then you add the regions by clicking once to start the region and then clicking again to finalize it. So it's not a dragging motion. It's really just one click and a second click to finish it. And you can draw however many you want. If you have a bit of a complex layout that is not just a simple block, you can also adjust this by clicking on let me zoom in on the region borders and then using these anchor points to adjust the region a bit better. You can also add new anchor points here and by just selecting it and dragging the layout region where you want it to be. So this way you can get some more exact layout depending uh what kind of complex shapes you have. Then the next thing you want to do is add tags. So the layout tags in transcribus are a label that you give to a specific part of the document structure and this describes what this is not what it says. So it's an indicator that helps you preserve information and context and you can add the tag by using your mouse and right clicking a region. And here you can see assign structure type and I can select one of the structure tags. For example, this is the newspaper. What do we have here? I select this region and this is shelf mark. If you notice that a tag is missing, for example, I want to let's add another text region here. I want to tag this um with the tag name, but I don't see the tag name appear here. What you can do is let me zoom in a little bit. Go to settings here in the bottom right corner. Open settings and then you can see an overview here of your structural tags and you can check those if the tag that you need is already available. We can see here. Okay. Name. I turn the toggle to enable name. And now once I click on assign structure type, I can see that name has appeared as a tag. And it already appeared here on the left in the image as well as on the right here as well. If the tag that you want to add is not available here anywhere, you can click on edit tags in collection setting. This opens your collection settings and here you can add new tags to your collection. So if you want to create a new tag, you click on create new. Now let's call this example webinar. Even choose color. just with turquoise. Um, and create the tag. Now, if I go back to my page, let's just reload the page real quick. Let's save our changes first. Thank you for the reminder. And then we should be able to see our new example tag. See if we can see it already. Our example webinar. Yes. And it is already enabled. So this is how you can add new tags to your uh document uh and to the documents that you're working with. Now one thing that you can also do is um just a small tip, you can also use doubleclick to open the structure tag menu. So if I enable this, instead of having to right click, I can just double click with my left mouse um button and the tag options are visible. So right click also opens some other options, some other settings and double click just immediately opens uh your tag overview. Now if you train your uh field model and because you can train your field model to only recognize the regions you need, you can also choose to train the model to only recognize these four regions, these four fields that I have drawn. You can choose to not uh draw other regions and to not have these uh recognized in the field models. So really you can see what information do you need from the documents you're working with and you can choose to leave out uh some information. Of course you can include it but just so that you know you don't have to um tag and mark everything. Then we save uh our changes. We already saved it for now. And you can also change the status of your pages. So we now change the status to ground truth. Um as mentioned before to train the model we need those accurate examples to show the AI what we wanted to learn and we call these accurate examples ground truth and with this status you can indicate that we have now created this accurate example and it can be used to train the model. Um again to start training a model we recommend at least 50 pages and more pages for difficult layouts or variations. You um can create a larger amount of training data in a quite simple way and that is by using a sample or creating a sample. So to create a sample, this helps create or yeah create a balanced set of training data of ground truth data. And this is especially useful if you have a larger variety of uh of different layouts. And to do this, you select the documents you want to include. We'll now use our example collection of the Peabody Peabody newspaper index cards. And then you go to our action button and you click on create sample. You can give it a name. We'll call this index sample. And we can see here can add more documents but we have already selected the one we want. And then you can select the number of pages you want to include. So there can be an absolute number. So for example 50 pages or also a percentage if we say 20% you can see 47 random pages will be selected. Now we can click on create sample and it should now create automatically a sample of this whole collection. And you can use this sample then for model training. And because uh if the sample is large enough, it should then add really randomly every type of uh layout of field layout that you have in your training data. uh for the model training. Now once your ground truth is ready, you can start with the model training and this is my cue to hand over to Sava as she will explain how to train the model. Thank you Elena. I'm just going to share my screen. Okay. So let's look now how to train our uh uh mod our first field model. As Helen said fields models can be trained to automatically recogn recognize and mark certain layout component of the document. uh when we have our uh at least 50 pages of ground root we can train our first uh field model. So we select the documents or the pages um with containing our ground. We go here to this drop-own menu and we select trade models and then field models. As you know we have it's possible to train four different type of of models in transcribus text recognition models baseline models fields and tables and today we focus on fields models. Um yeah we are just repeating that you need to before starting the actual training you need to create your grant route and 50 pages of training data is a good uh start as Helen said it really depends on the complexity of your material. Um if you have just two uh structural tags, two type of regions and the pages are very simple, for example, the the main text and the marginalia, uh 50 pages might be enough. Uh if the documents are uh more complex and you have uh 30 different types of uh regions just on one page because you might have very complex forms with a lot of different pieces of information. In that case, you need to increase the number of uh ground root pages to at least I would say 200 uh to 500 ground root uh pages for very complex layouts. The same with newspapers. If you want to train a field model just on one newspaper type, 50 or 100 u pages of ground truth can be enough if the newspaper layout is very homogeneous. But if you want to train a field model that can recognize different types of newspaper layout, you need to increase the number of ground pages. So just so you know the logic uh between the number of pages to to include and it's always good to do a test with uh to start with 50 pages train first field model like as a test and then increase uh the number of ground pages. Uh after you select train and model um the this the button we saw before um the interface ask you to se to confirm your training data. So all your ground pages or all the pages selected um go to the training data. Uh the training data is the pages the model is trained on. So here you just need to confirm that you want to train your model on those pages. Uh then the second page uh is the tag selection. You need to uh select to choose the tags your model should be able to recognize. So the training will happen only on the tags you select in this um in this page. Um, if you don't select a certain tag, the field model won't be able to recognize it, even if it's in your ground pages. Here, it's also possible to flag the option that Helena mentioned before, recognize untagged regions. Um so if you just want the field model to draw the regions in the way you want uh without adding tags u you can do that by flagging this option. Uh and this is what the baroness of blocks public model does. So it recognize the blocks the the regions um without adding any tags. And here we also have another option option called train online polygons. Um this is a diff a slightly different type of field model. Um if you select this option the field model isn't trained on the tech on the regions and on the tags but on the line polygons. The line polygons are uh the polygons encasing the text uh the lines of text. Uh sometimes with very special uh documents um the auto the the automatic computing computing of the line polygons isn't satisfactory. So you see here we have this page where we have or we have very thin uh and long line polygons. Um if you just run a a very simple text recognition uh on those pages the risk is that uh many words aren't recognized not because the text recognition model is bad but because the line polygons aren't accurate. In such cases, you can train a a a field model on line polygons to get the polygons correctly recognized. But this is a very rare case. So in with most with many documents, I would say 98% of the documents or 99% of the documents, you don't need to do that. Uh but if you have very large um characters or uh your documents are very peculiar, you might think about train using this option. But now look, let's go back to the standard fields models. Uh on the next >> sorry can I just ask sorry uh there was one question um about the training of the field model and it is does it matter for the training whether there's already recognized text in a document or do I have to train the field model before using a text recognition? I think that fits in here quite well. >> Yeah, thank you. Uh no, it doesn't matter if there is text uh on the document. you can create the ground root. Uh it it just looks at the regions and the text. Uh if there are baselines or text or recognized text, it doesn't matter to to the training. And then we have this last uh this page where you are asked to select your validation data. automatically transcribus uh prompts you to select 10% of your uh training data. So during the training 90% of your grant pages are used to train the model and 10% are set aside to test the accuracy of the model. Um it's so we recommend to stick with the 10% uh to to do the 10%. And then we have the model setup page. Here you need to add name, description and further information about your model. And here also possible to um change the advanced settings um the trainings. Uh these are the four advanced settings that you can uh change. Uh for the training cycles and the learning rate, we recommend to stick with the default settings, the recommended ones. Uh the training cycles are the number of times the model goes through the entire training data set. Uh by default, I think is 15,000 training cycles. Um, you can increase this number up to 30,000. Um, and I would do that only if you have many ground pages. Let's say if you have a more than 500 ground pages, it might make sense to uh increase the training cycles. If you have a very limited number of grant root pages and you increase the training cycles, the risk you the risk is that the model overfits which means that it learns very well the training data but then it doesn't um learn to generalize on new data. So there are also there is also risk if you increase it but you don't have enough training data and then you can also choose between two different type of uh backbone architecture. Uh one is standard and one is enhanced. Uh select the standard one if you have simple document structures uh like the index card we show you before. uh select the enhance one uh if you want a field model that can recognize different type of layout at once. For example, if you have a a form that changes over time, uh you can train one field model on different uh on these different forms. U you just need to increase the training data to make sure that all the different type of forms are seen in the training. And it could also help to select the enhance um architecture. When the training is uh finished uh we uh want to review our model and the value that uh express the precision of our model the accuracy is called mean average pre precision and you can see it here in the models card. Uh this value is a complex me measure that evaluates how accurately the system detects text regions considering whether they were detected and how well their size and shape match the validation data. If the mean average precision is over 60% um it means that the model is uh quite good and usually it delivers satisfactory result. So don't aim to reach a 100% uh% accuracy because it will be very difficult uh I would say the impossible um also because uh it also depends how this value is is measured. So it will be impossible to reach 100% accuracy. uh but uh if it's already over 60% it means that it's good and very likely it will give it it will produce the out outcome that you expect and always run a recognition test with the model on a few pages to evaluate the results yourself because even if the uh mean average precision isn't that high high as you would have expected fact it could be very good during the recognition and you don't need to and maybe you don't need to increase the ground and train a new version of your model. Uh you can also in the model card you can also check the uh number of the instances per tag to understand how often a model has seen one instance of a tag. Um, so if you you can check these numbers and if you notice that one type of tag is less represented than the others, you can create some new ground truth pages including uh more of those of of the pages with with this tag and then retrain a new version of the model. As I said, it's always possible to train multiple version of your model. And when you have the first version, you can use it to speed up the creation of more ground pages. Because if you have a model uh to start from, you can recognize new pages and just correct the regions instead of drawing everything by by hand. uh and uh in the end uh you can train a a new model with the old ground root and the new ground pages. And now let's look how it looks like in the the platform. So I have uh this ground truth pages. Uh I and you see here the banner at the top is dark green which is the color for run in transcribus. I select my 51 run pages. I click train model build model and automatically all the pages I selected are assigned to the training data. uh I go to the next page uh and here I'm asked I need to select the tags I want my model to be trained on. So in this case is details um let me me check name newspaper reference and shelf marker. uh I can also combine it with recognize untagged regions. Uh what is not possible is to combine it with uh the training on line polygons. So either you train the model on regions or you you train the model on region or you train the model on line polygons. U it's not possible to do the same uh both task at the same uh time with the same model. Um here uh I have my validation data. I can also change the percentage or manually select my validation data but it's safer to select it to keep the automatic selection. uh because in this way we know that uh the validation data is u has been randomly selected uh and so it's unbiased and uh it should give us a realistic value in the accuracy and uh here we can add the model name the description here we have our advanced settings for now we stay with the standard option because this is simple model. We go to next. We review all our data and we can start the the rec uh the training and we when we go to the training lab uh we see that the fields model training has been created. Uh we are first in the queue. Uh so we need to wait uh probably a few hours uh before the training is completed but we will receive an email when the training is uh is done. And when the training is finished, we can go to uh under the models category here uh select fields and we see our uh the models we have trained. Uh we already trained the model on those index cards. Uh you see we have 47 training pa pages in the validation data five in no uh 47 in the training and five in the validation and the mean average precision is 88.77% which is very good and also here if I click show details I can see the number of instances for each tag after the training uh let's see what we do next. So how to use the fields models? Um >> can I s >> Yes. >> Can I just quickly um ask two questions before we move on to using the model? One question was about the training process and the question was if you train the model on line polygons can you skip the layout recognition so the baselines recognition later uh yes you you need to do that um because if you run the in that case you need to run the field recognition and then directly the text recognition uh remembering to flag the option keep existing line polygons. So if you have your um field model train on line polygons, you use it for the recognition of the line polygons and then you go directly to the text recognition. Uh but please remember to flag this option otherwise transcribes deletes the line polygons and uh draw draws new ones. >> And the second question was about retraining a model and maybe you could just quickly reexlain the process. The question was when training a new model after the first training do you use a public model or the previous model trained on the same document? So I assume that the question is about what the retraining is based on. So many uh we have some public models for fields um but uh probably they if you have very specific uh type of layouts they won't work on your u documents. So the strategy here is to train a first version of your field model uh on a few pages, let's say 50 pages um to have a sort of a draft model. Uh so instead of creating um 100 or 500 grroot pages manually, uh you can start small. So with 50 pages uh you create 50 pages of ground root. You train your first version and then you use it to recognize uh the remaining 450 pages. Let's imagine it's a very you want to train a very big model because you have complex documents. Um, that way on those additional 450 pages, you don't need to draw and tag everything manually, but you already have a a pre-recision done by the first version of your model. And this should speed up the creation of your ground root because you don't need to draw manual everything manually, but you need just to adjust the the regions because maybe on some pages they are too short uh or too too big. Um and when you have checked and correct all those 450 pages, you start a new training including the early 50 pages of ground root and the additional the new 450 pages and you train your final uh field model. With field models it isn't possible to um select a base model as we have we have them with text recognition models but this is not possible to to to do with field fields models. So uh you cannot train a field model on the top of another one. You always need to select all the bas all the ground. I hope I answer. Thank you Sarah. >> Yeah. So now let's see how to use fields models. We can directly look at it inside the the the platform. So let's go back to our collection and let's take uh one page. Uh we are here. We have our index cards. Now I'm showing you how to do that for one page. What we recommend is to test it on five 10 pages and then when you have found the correct settings the satis that gives you a satisfactory result you start the recognition for the entire documents or the entire collection in a batch. So you don't need to do that page to to run the recognition on each page separately. Um we are here and we click process with AI. The first step is the field recognition with the model we trained. Uh we select a field here. Uh and our model call in the card model 2. Um and we start the recognition. Here we had have the advanced settings and we will look at them in a minute. So let's first start let's first start recognition. It's running. Okay, now we have our region with the correct tags. Uh what we can change in the advanced settings. So one setting we can change is the detection confidence level. So how confident the model should uh should be. uh it's what uh Helena show before with the newspaper. So we are telling transcribus um to to recognize um the regions with a certain confidence. So if in this case if um trans if transcribus if the confidence is less than 47% uh transcribus won't show us those regions. So here when it's set to 75% we have two people recognized here. uh when we lower the confidence level we have three of them because the model wasn't that sure that this should be recognized as a person. Um so if you increase the confidence level uh it's likely that you will get less regions uh but they are more accurate. If you decrease it you get more region with the risk that some of them are wrong. It really depends on your material. So you can do a few tests. Uh the default value is 0.75% which works well on most of the cases. Then you can uh change uh the shape detail levels level. So how you want uh how you want the shape of your regions regions. If you want just a rectangular shape or a polygon and also this is medium. So you have a uh a polygon which is however close to a rectangular and then if you set the shape detail level too high it's more um it follow more the the content uh of the region. But remember the shape of the region depends on what you um had in your training data. So if in your training data you created just simple shapes rectangular region even when you set this value to high transcribus will create s simple shapes. uh then there is the possibility to add the field recognition to exist to an existing layout. Let's imagine you have this document and you already transcribe the main text and now you want a model that recognize and text the marginelia but you want to keep the main uh the other text you need to flag this option and this option is also important if you want to combine tables recognition and fields recognition. So if you have a table but also a text region on the page and you want to combine the two of them, you first need to run the the table recognition. When the table recognition is finished, run the field recognition flagging this option and transcribus will keep the table and add the the regions. Uh and then we have a option that we recently introduced. It's called merge over overlapping regions. And also here you have a threshold. Uh if two regions with the same tag are uh overlaps for more than this threshold. So in this case for more than 65% transcribes automatically merge them. Uh but please remember that they should have the same tag. Uh and also that this option doesn't work if the shape detail level is set to high. Uh so it works only with the detail level set to low or medium. And then the next step and we can look at it here. Uh we have our regions and now we want to find the lines inside those uh those regions. Uh the next step is to run the layout recognition and uh select one of our public models. Uh the the best ones are mix line orientation, universal lines or horizontal line orientation. And uh here remember to modify the advanced settings. In particular, you need to you must select this option keep existing text regions because we want to find the lines inside our fields. We don't want transcribers to recognize other regions. And it might also be helpful to decrease the minim minimal benzlank length to low especially if you have uh very short lines like digits. Uh and also to flag this option split lines or region border. U so the line ends with the the region and we can now start the layout recognition. And the next step when we have our lines, the next step is the the text recognition. Okay, it's done. And we have our regions and inside the the regions the our lines. And then as I said the last step is the text recognition. We can select a super model or in this case transcribus print M1 is enough because the it's typewritten uh text and I can already show you the the result. So the final results will look like this. Now we want to the final step is uh exporting our um fields. Uh we select the pages, we go there, we click on export and the best option um to extract information uh extract the data uh when you have a fields model is to select the spreadsheet file format and here the type you want you need to select is a structural structural regions export. This will create a a spreadsheet. Uh each row is a page uh of your document and each column is a a field. So the header contains the structural tags. Uh while in the cells you can you the cells contains um the the text recognized in that specific uh region. And now I leave the floor to to Helen for the last part. Thank you, Sara. I'll just share my screen real quick. I think it's lagging a bit. Let's see. All right, we're already at the end of our webinar. So, I'll just be doing a short wrap-up of what we've seen today and what we learned. We go. Think you should be able to see my screen. Yes. Nice. There we are. All right. So, we had a look at a few things. We had a look at how you can prepare the training data for your field model, how to train your field model, and how to evaluate it. And we also learned how you can retrain and improve it, and of course, apply the field model to your document. And the whole idea with transcribus is that we really want to help you save time with your work with historical documents and provide a tool that can be learned quite quickly how to use or you can learn quite quickly how to use it because we think that AI should be able to be understood and also to be learned and also controlled and that technology should be for everyone and not just tech experts. So we really hope to create tools that help uh you work with historical documents, with handwritten documents, uh save you time, save you work. Um because it's not just about reading text while it is a big part is reading the text, but also understand it and uncover historical information. And that is why we really want to provide the best tools to make history accessible um and provide the best tools for you as our users to help you work with your historical documents. Um but who are we? We are REIT co-op. Uh transcribus itself comes from the university sector. We started as the REIT project at the University of Insburg and we are now a cooperative with over 250 members public and private institutions and individuals. Our core idea is the further development of the transcribus platform and our motto is really purpose over profit. So everything uh in terms of uh profit that we make is reinvested into transcribus and distribution is excluded by our statutes from our cooperative. So everything that we put back into transcribus is really there to develop new tools, new features and implement them. for example, such as the field models which are uh I think we have had them for one and a half years now and uh also uh table models which we had the webinar last week text or image matching and also one of our new um uh features that is coming this year which are the endtoend models. So with the field models, we showed you how to h how you have to do the layout recognition and the line recognition, the text recognition in three three different and separate steps. And we have been working on a solution to and not have to do these things in separate steps and end to end models will be able to take care of all these three steps in one. So layout, text, and named entity recognition will all be able to uh be done in one step. We can't say too much now, but this is something that will be coming later this year. Um, and we invite you to stay updated by following us on our socials or of course uh take part in one of our future webinars. Um, but yeah, this is something that we have been working on for a while and we're excited to to be able to show you later this year. As you can see, we have a lot of amazing members in our cooperative who are all very involved in providing feedback and involved in furthering the technology and the community also and really creating better transcriptus tools for our users. And now, yeah, now it's your turn if you haven't already. I think most of you uh have indicated that you have tried before. if you haven't yet, it's your turn. Um, also to try training field models. We really hope that this was a helpful instruction instructional webinar to training field models. Um, as mentioned before, we'll also upload it on our YouTube channel. We have more helpful information on our help center. Here's the link as well. Um, you can find more instructions, helpful uh posts, step-by-step instructions on how to work with transcribers, how to train models, and anything that you can think of is in there. Uh, we also have uh always some upcoming webinars. You can find them on our events page on the website. You can also find our past events there. I always try to include link to the webinar recordings in the webinar description. So, at least the webinars from this last fall should have the webinar recordings also linked. But if you don't want to click through all the past events, you can just go to YouTube and rewatch our latest webinars um and click through those. The nice thing is you can just pause, try for yourself uh in your account in transcribus and then continue watching. So that's also very nice and that is why we upload them there. Um yeah, you can join the conversation, follow us on our social media accounts. We post regular regularly also updates on new models, new features, um events that we're attending if maybe we can meet in person somewhere. And that's how we uh want to stay and try to stay connected with our user community as well. And I think that's it for now. It's 5:02, so we're quite on time. If there aren't any more questions, do we have uh some questions that we should still address Sara or >> There is one about smart extract models. Uh if >> yes, >> we will announce something this year or when >> yes, so our end toend models, we will be announcing something this year. Again, I can't say too much for now, but we can say that they are coming and that we've been working on them for a while. Um, but we will update you with information as soon as we can and we will be excited to do so once the time is here. So, I think we'll just wrap up the field model webinar for today. Um, if you have any more questions that we weren't able to answer, you can always also reach out to our help desk. Um, we also have a great team there and they'll do their best to answer your questions uh via email. All right, then. Thank you so much for joining today. I hope you'll all have a nice evening, nice morning depending from where you join uh our webinar and then I can just say thank you so much and I'll hope I'll see you next time again in one of our webinars.