 Session three in the program now deals with what is new in transcribers and we will be hearing about the likes, trends, thoughts, transcribers, that the technology is offering. First speaker up on the stage is Gondram Leifert from Euro University of Rostock and he will be telling you all the neat and nice details of what a 100-compact recognition costs. Yeah, I will talk about HCR Plus behind the transcribers. HCR Plus is developed at the University of Rostock in the SITLAB team together with Tobias Bruging and some other guys from our SITLAB team. Well, you have what to talk about. I want to talk about some technical parts, probably not that interesting for the most of you. Then I will make a comparison between the old and the new HCR. What can you expect from the new HCR? Then I will tell you something about common errors that you can make when you transcribe and trade in your network, which can make the HCR working in Rostock. Yeah, then I'm ready for a discussion. So, first of all, what is changed between HCR and HCR Plus? Not much for you. The input is still an image and the output is the transcribed text. The whole thing is trained with connection and temporal classifications and algorithm. Yeah, it works very well. What did we change? We changed the software. We had our own software and now we use software based on TensorFlow from Google. We changed the hardware. We were trying on CPUs. Now we're trying on graphical processor units, which are a lot faster. The architecture of the network changed. They became deeper and larger. And also the pre-process we have to change a bit because it has to be faster for the fast networks. Yes, here you can see on the left side the input is always an image and the output is not only a text, it's also the probability of a character at each position over time or from left to right. So, this here, black means 100%, white means 0%. This so-called confidence matrix or confmat contains information what is written in the image. So, this line here, this red here, means that the 3 is wrapped by the neural network to nearly 100%. Here the network is reading the point, here's the space and so on. So, the system is more or less working. On the right-hand side we see the structure of the HDR. It has some layers. And after each layer, the presentation of what is saved in the image is changed, more and more abstract. The input is the image and after each of these representation layers, the representations change until to the end we have this confidence matrix. And with the help of this confidence matrix, we can produce a transcription and we can also use this confidence matrix for the keyboard spotting search. So, this was the technical stuff. Now, what can you expect when you already had a train model in Transcubos? Yeah, we had the data set. I realized there's a typo I get. I guess it has to be a T. The old HDR, this train on a data set, the result was on the validation set, on the test set, it has 21% character error rate. The new HDR in this case comes down to 9.86%, so we can reduce the error by 50%. The time spent for training was not 20 hours like before. Oh, it's 5 to 6 hours in this case. And at the same time, we can train with 8 times more lines. So, the training got lost faster. The same for other collections like from AVP, we have the same. It is again faster and we come down from 19 to 6.35%. The best thing is Conceals protocol of Dr. Albaugh that we come down from 10% at the beginning to 2.6%. So, in this case, we nearly enhance the system by 75% character error rate reduction. I think the HDR Plus is not available for all of you so far because we just have one computer and so many people who want to train. So, you have to understand that we can open it just time after time for everyone. Now to something for you, when you make a training, you have to produce training data. We already said, the more training data the better and you already know that. But another big issue is also the quality of the ground shoes. The perfect thing would be 100% correct lower errors but also you make errors, not only the HDR. When you transcribe, there is more than one people and you have to transcribe. You have to define some rules how to transcribe, how to handle minus between words Should it be separated by spaces or not? Do you have an extra character for hyphenations or do you use a minus or an equal sign? So, if you transcribe with more people, tell them how they have to transcribe. Here's an example. The long S in fracture is, you want to transcribe it as normal S or long S. Please make one of these versions. Either the normal S, this is not the problem for the HDR to say the same S. It's this position and this position. You can also make this but please do not mix it. You know that one cannot know why it has to recognize long S here and not long S here when they cannot read. So, the ground shoes was this year. I found it in transcripts, this was the ground shoes. But you need words to tell me how to transcribe. This is very crucial otherwise the network is also confused. It's the same for abbreviations or diplomatic text. Often the goal you want to have is modernized transcripts like expanded, abbreviation and so on. This does not work very good. So, the network more or less can only read what is on the image. So, do not expand abbreviations like conclusion or file touch trigger. Just transcribe the short path. When there's just one character between like when you have the ID Das and the middle character is missing. This is possible. The network can maybe transcribe one more character in a row or two more character but it cannot expand the abbreviation. When you produce ground shoes there are in transcripts several steps what can go wrong. For example, typical way you transcribe a page then you replace HDR to check how good HDR is working. And then you start a training. Then maybe you start a training with the HDR output. This is there. Not a good idea because the HDR makes errors. So, if you want to train you have to check that the newest version of the file which you can find when you click here in the Graphical User interface. This video is shown. You have to check that the newest version is not a version produced by an HDR. Otherwise, this is used for training. And the best way is to tag pages as ground shoes. This is possible at the top of the Graphical User interface. Here you can select from new to the process and to ground shoes. And afterwards, when you start the training, you can check this button. You use only ground shoes version for training. This checkbox is available on the training window. And when you choose this, you only take these pages where the status is set to ground shoes on the page. So you can be sure that no HDR result goes into your training only the ground shoes pages. Then sometimes the HDR phase. There is one thing you can do. At the top of the Graphical User interface you can show some surrounding polygons. This is a bit complicated to explain. The input of your network of the HDR is a line surrounded by the polygons. This polygons is calculated by algorithm of us. And you do not see them in your Graphical User interface. You have to select the check button and make it visible. And when you make that, sometimes you see that they make errors, especially at the borders of regions at the top or at the bottom. When they surround two lines, for example like here, then the input of the HDR will be two lines. The neural network is totally confused and will not transcribe anything. So if you see some things put on the polygons, this is the reason why your HDR fails. But this is our to-do list. We will try to improve this polygo method. And hopefully it will be no more a problem in two or three months. So the question is, thank you very much. Question now. My phone is coming. Thank you very much for your presentation. My question would be, I'm new to this. Will this HDR plus? Is it also an open source? Can it be integrated to other projects than transcribous? No, it's not open source. But there is the REST API that maybe is possible to use it without the user interface. You can make a request, I guess, at a server and say, use this neural network on this page. And give me the XML result. And why it is not open source? Because it's funded by public money or not? It is funded by public and we have the open access things. But some of the tools which were before, which were REST access, before the project started, does not have to be made open source. So yes, when for example someone comes in, this is how it is. So if you have something before the project started, three years ago, the software or parts, they do not have to be made open source and they are still used to these things. For the questions? A simple technical question. Will it be possible in the future to launch two models in the same page? Because we have sometimes a page with two and writing text different from two hands different. And I guess if it's possible to have two models for something or a combined model. So it's really we have some transceivers interfaces for HDR. And therefore, for example, transceivers supports regions and you can have different reasons. And then you can say use this model for this region and this model for the other region. This is possible in the interfaces but not so far usable via transceivers user interface. So it is already possible but in the user interface not so far. One comment from my side concerning the usability of HDR Plus. As you have seen, we have a new release came out this week. And with the new release you can already use HDR Plus models if they are there and if they are opened up to everyone. The training of the HDR Plus model is working as well. But here we really have to comment that if everyone now is interested, press the button. Our server is from Cache. It's just one server. It is a powerful GPU server but just eight jobs can be done at once. So it takes a while. And therefore we came up with the idea actually that we will retrain all models which are there in an automated way. So this means that we will start with the latest ones and more updated models will appear simply in your collection. And in that way we hope that it can be done... Utilized to serve as the best way because always eight jobs are running and we care about this automatically. So that's the way we decided to do it. And then when the first spy for me is over, then we will unlock the HDR Plus training of course also to you. Thank you. I have a question about the abbreviation. Because you said that the abbreviations doesn't work. I mean it cannot read the abbreviation. But yesterday I heard from Ginter I think that it does. So I would like to... Because I don't know exactly how the computer read the text. So I am curious how it is finally. I mean is it worth to write abbreviation if they are easy let's say or totally not. Only the letters which we can see. Because for example let's say that we have some abbreviation which is always the same three letters with the line up. And so we should write for example that in many of our Christos is X per H never mind. But is it worth to write it like it should be or just these three letters? So there's just one letter missing. Or you expand one or two letters to two or three that's fine. But assuming our algorithm works like that. You can say every of these slots here. You can say one character. So assuming you are training with 3rd January. Where should you read the last five characters here in this part here? There's no space to read characters. But when you just want to have one more character the system can put it in here somewhere in the slots. Like yeah when you have some. What do you think about this HDR plus? It's the same for both. It is the same problem for both. So when you have just one character or so that's fine. This is what good can be. But when you have DR for Doctor or other longer expansions then the network will fail. It will try and it will probably match two or three characters but not the old one. And the character error rate will be there. A little bit possible in the future. That I make some rules in the transcription. For example I have XPS for Christmas. And then I say to the training software every time you are reading XPS. And the transcription is XPS please. Yeah make the right it like Christmas in it. That every user makes a rule for it. These things are a lot of work because you have to upload the dictionary. Which have a kind of standard. And then you can apply them. I'm not sure this is how high is it on the priority list. It's really possible yes. Same with abbreviation there. Or with name entities it's the same. Okay let's thank Gondram again. In this really interesting session is our read colleague Erwe Bichard from Navelabs in Grenoval. And he will be telling you about tables and about document understanding. Thank you. This is common work with Petra and Francivia and two colleagues from Navelabs. So just to give you an anecdote. The first time I saw similar images from it. My reaction was not so enthusiastic. And I was really wondering if we can do something for such a meet. One two years later I would say we have some more. So here what I'm going to talk about. So first I would like to define what I understand as table understanding. Just to make sure you are on the same page. And then I'm going to go through several scenarios. Which are strictly the way the table processing a part of the read project has evolved. And then I will go through a couple of examples to show you what we can do. And then I will discuss an initiative we are going to physically build open source data sets for handwritten tables. That we can share with the academic community. And so that everyone is able to tackle the task and share their results. And then quick speech or mixture. Here is a typical table we have from the pastoral collection. And so by basically table understanding. What I understand is basically the recognition of the row structure and the column structure of the table. So the first task is to spot the table. And then to recognize the rows and the columns to get the cells. You can add extra information to the column meters or row meters. And here are textual elements outside the table. And so that then you can automatically access the row structure of the table or column structure of the table depending on your data. So this is what I'm going to talk and how we do this. So this is how we do this. So scenario zero or I would say even minus one. Basically suppose you have just one page with one table. So what you can do now is to use the expert user interface where you can manually design the table yourself. And then you will be able to export this table into an Excel spreadsheet if you want. So this is currently viable. How to do this? There is an how to return by effort that you can download. The second scenario is basically you don't have just one page but you have two. But it's a kind of creative book where for each page you have exactly the same table. Very common in current times. Here the solution designed by Florian in the room. But basically so we add the user through the first page of the document. And then the two will be able to match this template against the remaining pages of the book. And then you will get the book for the structure in terms of what you will get with this tool. This is the column structure of the tables. Because the template in terms of columns is similar from page to page but not in terms of rules. So it's currently viable not through the graphic user API but through the programmatic one, the REST API. So you need to be familiar with this API if you want to use it. Or if you want you can also download the tool and apply the tool on your images. So it's explained in the same how to write about the table. The next scenario is you don't have just one book but you have hundreds of books. And each book will have a different template. So you don't want to take through an each book through a table for the first page and then process it. So the purpose here is to automate this. And the tool we are currently developing is basically a tool which takes a book. Kind of mind the set of pages in order to automatically generate columns to control the table. And then apply the previous template matching tool on these tables. So you can get a column structure for each page. And so you can do it automatically for a large number of books. The quick example here is from the same book you have three different pages. The same table. And basically the tool is able to recognize the vertical organization of the tables. So from this vertical organization what is missing is the horizontal one. The row one. For this we use basically input. So we recognize the text line of the table. So if you apply the level to integrating the tables you get automatically this. As mentioned several times but it's worth mentioning again how the tool developed by Tobias is really, really good. I've been working with this for more than a year and I'm still trying to find page variables which will be changing. So from this set of text lines what we can do is to categorize each text lines with many good information. For instance here the tool is able to recognize elements outside the table and also the column header automatically. And if you want really the row structure the idea we had was basically apply to automatically recognize the text line which are in yellow. And thank you. And which basically starts those text lines in yellow basically corresponds to the first text line in the given cell. So in the row. And with this information we are basically able to generate this kind of output. So here we have the full table structure with the columns, the rows. So from this we generate the cell structure. The headers are computed and elements outside the table are also recognized. So this kind of outputs is currently possible if you know the table from the given book. So sorry the template. So here the template was provided. So the template finding the column for the table was provided manually. And then the row recognition was automatically done. Depending on the tables you have you can also try various over possibilities. So here the graphical lines in the page were recognized. And the idea is basically that from this information you should also be able to rebuild the table structure automatically. The issue is that sometimes you don't have any graphical information in the table so you cannot only realize some graphical information. You have to combine both. So the textual information and the graphical line. This is currently what we are trying to do to merge all this type of information together. So all the methods we are using are based on machine learning. So for this, for the HDR, you need to provide a system with samples, annotated tables. And so we have such a dataset for Passao, for Math as well. And in here we are building a more generic dataset around the table. So we did ask you a couple of months ago to provide us with tables. So we had around 25 different providers. So now the dataset is composed of roughly 1500 images from very various tables. Mostly handwritten, we have also some printed tables as well. And so we annotated the tables. So first we had to write guidelines. It was not always easy when you see a table to know how you would like it to be structured. Sometimes you have different tables in one image. Very similar, you know, if you want one tables or two, it really depends. And so the dataset will be released for the competition next year. And the one question is basically, can we with this kind of data generate kind of generic model for tables which is as good as the text line detection we had in the project. So we'll see. I guess one backup solution is for given collection to annotate samples of the collection and just to train the system on this specific collection. So I don't know if we'll be able to have a kind of generic model for table processing we'll see next year. So yeah, next year. So a schedule is a bit fuzzy, but this is the best I can do. So what we want to do is to release a prototype for table processing, I would say by the end of the year. So it will be available first as standalone to it's written back then. So you will have to download it and turn it on your on your connections. And I get the plan is to integrate this into the transcript server next year. And so Victor is the main conference in the domain of document analysis and recognition will be next year in September. And for this conference we will organize a table recognition competition, which is a data set which will be valuable beginning of next year I guess. So if you weren't interested in this work, we published with a fan for a paper on the last task workshop this year, about the work done for the Pascal collection. I will try to describe as much as I can in the variable we need to write for the read project by the end of this year. If you're interested, the current version of the software is available through GitHub. Don't try to use it currently with couple of weeks at least. And we have already some data which was used for the article and from the Pascal collection. And those data set is publicly available on Xenodoo. So I don't remember the size of it, that's 200 or 300 pages of the Pascal table. It's completely annotated. And if you have questions. Questions on tables. Shape them up now or never. This is your chance. Don't you have to tell the system which way to read it? You read it from top to bottom or you read it from left to right? Don't you have to sort of add this before you scan or transcribe your documents? Well, I mean, if you have a table, you can read it from top to bottom or you can read it from left to right. The structure of the table will be two dimensional structure. So basically at the end of the process, you're able to do both. It's like in Excel when you will be able to navigate column-wise or row-wise. That's the purpose of the task. And then it depends on the data you have in the table. Sometimes there's the written. The row reading will make sense. Sometimes it's the column reading. It depends on the task. So it's throughout the period. This is not the end of the process. Once you have the table, then if you want to basically feed the database with the information in the table, you have still to do some information extraction process and normalization and a couple of things. But at least you have this two dimensional structure and you can use it. Thank you. Could you say something a bit more about the content? The content at the bottom will be maybe nominal data or dates. So this is what we have done for Basel. Here it will be true in your case as well. You need basically to know, for instance, that in a given column you will find specific information. And then you have to design an information extraction tool usually dedicated to your task in order to extract the information automatically. Here I don't think that's a generic tool. I mean, you can have a generic tool to recognize dates or names. But if you have just one name in one column with a date, then you can put some semantic. You know, it's a birth date and it's fine for you. And that's how in that course, you have the date, but also the burial date. You have to know which one will be the burial and the date. And for names as well, in the marriage readings, you will add the witness as well. For a given reading, you can have four or five names. And then you have to find out the right one. So yeah, information extraction in current is not a generic. Thank you very much for your presentation. Your module, will that be able to use your model of the one you've been developing? I'll say like OCR program as well. We have lots of tables that are printed text. Yeah, for us it doesn't matter if it's handwritten or printed. I would be interested to hear what you are saying about how far it will be a generic tool. Actually, you got the point right in the first talk, if it depends on the data. And actually I think the success of the current layout analysis is based on the thoughts of the BNR team collecting the read that data set, which seems to be very, very representative. That you are probably finding any document that's present, really hard. And with the data set you are already, you are collecting right now. Actually that's a really important step to improve the baseline detection for this table. Now for the table scenario, and to cluster these for the entire, to get the entire complete structure of the table. I think it's not so far away. So basically if you got all the baselines correct, then I think you don't have so much problems to get the root lines, right? Or your tool is working quite well right now. Sometimes you don't have, what do you mean by root lines? The horizontal boundaries between the rows, parameters depending. And when you see the different kind of tables in the data set, you are not at that point currently. And we'll see. It's always hard to say. I would guess it's within the next year based on the data set. If it's so representative and it can work with it, maybe I don't find the word representative because the feature type is there. And therefore it's cool and it's necessary that these algorithms as well are trainable and are adaptable to the needs of the users and the different tasks. And another to avoid this annotated data problem would be a frame. Frame now is to use sensitive data. Basically we are able to automatically generate tables which are annotated. And then we can train the system with this sensitive data. So if you know from all the time of the tables you have to process, you can generate tables which are very similar. And then the first results are running correctly. So you just have to give some parameters to the generator and then this is the way you can generate the data here. So we'll see if it's focused now. Well, thank you very much, Avi. I think there's a lot more things coming up on tables and on HDR. So let's give Avi a hand. Pleasure to welcome to the stage our head marketing person for Keeber's Backup. I'm going to move back, we'll also be speaking about sharing HDR models and about the new fancy interface the transfer was back up. Okay, so this was a good introduction that I'm marketing Keeber's Backup and you will hear more about Keeber's Backup tomorrow morning and also see the nice implementation of Keeber's Backup application for the band time project. So, and this is the best marketing you can think about it. But actually I'm now telling this story about Keeber's Backup since a year or more. And interestingly the reaction of the audience is often more interesting but then it stops actually. And so I thought I give you a quick one more try and also going to show you what was integrated already in Keeber's Backup. And then I would like to talk quickly about with you what to do with the many HDR models we have now in Keeber's Backup and how to enable sharing. And also quickly I'll look to turns Keeber's back into things to start with today already. And I think you will also experience it in the workshop so I won't be very brief about this. Yeah, Keeber spotting. Usually searching is performed via a good transcription but for large collections also this HDR plus we will have the situation that you need a lot of training paper that you can process millions of pages. And if there are hands we've never seen by any chance the recognition rate will be not as good that the usual full-text search will produce good results. So that's a matter of fact I would say. And in many cases you will not have the resources through tens of thousands of pages and also a case where the recognition rate for unseen hands will not be so good. And here comes the idea of Keeber spotting. Is it possible to search not in a transcription but directly in the image? And this is the main idea of several approaches actually. It sounds a bit like magic. How can I find a world which is not actually recognized by the engine? And therefore I created a little presentation which tries to visualize this a little bit. And let's assume you have dogs, cats and birds and we want to train a neural network recognizing these nice animals. So we would label some images that there's a dog on this image and a cat on this image and a bird on the third image. We expect that the neural network learns and outputs this in a correct way. And that would be, I guess, a typical student work at the university in your course. And other applications, of course, recognizing images which have not been seen by the computer, by the neural network before. So the question is what is on this image? And this magic box is working and produces, as we have heard, a confidence rate. And in that case it says something between null and one, 0.6, and let's assume it says this is a cat. So this is the correct answer. This is something in between a dog and a cat. I was very happy when I found it. So let's assume it's not a bird, maybe a cat and this confidence of 0.6 is a dog. So the highest confidence level is the cat, but the correct answer is the dog. Okay, so if we transform this to the transcription and searching, show the transcription is incorrect. Because the system would say my highest value is cat and that's incorrect. But if you are now searching, I will say give me all images which are above a certain threshold. And it would return this image as a dog. If you are now searching for dogs, it would correctly give you this image because it's above a certain threshold. That's the main idea I think of figure spotting and the transcription still would be wrong, but your search results would appear on a relatively high rank. That's the main idea and this technology is now using the alphabet. So cat, dogs, birds are the single characters in your text or in your text. And the transcription are the characters with the highest confidence, but keywords searching provides you the string which is above a given confidence. That's again the magic of this technology which makes it so powerful and is not based on the text. It is based on the kind of indexing of the image. In Transcritos, we have now implemented this in two ways actually. One comes from Vostok, from the Citadel, one comes from Spain, Alejandro and from the Greeks. One uses the confidence matrix, so this what you have seen before, this character based equivalences. Or both are using this in a different way actually. So that is the implemented keywords spotting in Transcritos. You can direct the search after you have performed the recognition. This means that which is a very nice thing because you just run the recognition and you can immediately search afterwards. The way it is done is that each confidence matrix is open and this dynamic programming, the best matching strings and so on are formed. This is fast, but for large collections it is probably not really fast or it requires a strong computing power behind it. The other way comes from the team in Valencia here a lot of work goes in creating an index. So the index is a reduced confidence matrix based on some highly sophisticated processing. In former times they used words, now they are using n-grams as far as I understood. Therefore, the auto-propeller rate problem which always appears with indexes can be avoided. The good thing is that we have a really quick response time and we can use standard full-texted engines for implementing this. It looks like the interface is more or less always the same. It is not a final interface, it is really more for demo. This is maybe also one of the reasons why we are not working that much with student spotting. The user statistics and they are always very low. So this was one of the reasons why I wanted to mention this here as well. In this upcoming project with the National Archives of India we will implement student spotting and we will thinking goes in a direction that how to use this kind and how to design the user interface and currently we believe that it will look similar to what I have shown you here. So on the one hand you have a tree level which is for my point of view very important for archives because in archives what commands are typically ordered in a hierarchical way. The users which are searching for a word will get an answer on how many hits are on a specific hierarchical level starting with level one and then we will come to the sub-collection and so on. Then you can likely order the snippets or the hits in the usual way and on the right hand side we will have some facets like narrow research according to the document to the archive unit or to places appearing on the same page or to keywords appearing on the same page as you were looking for. And in the center of the interface I would imagine an image snippet which shows two or three lines so that we get a kind of context where we already have a first look at the hit at the result and decide, and I think this is very important, decide if this is relevant for you or not. Because this is something I'm currently missing very much with all kinds of user interfaces in the archival and library world that you stop researching until it's very seldom possible to export and to take the data with you. And that's, I think, is really important. So one other idea is that this is really just the first idea is that if you have maybe already decided that this is interesting for you you might even add a command or a tag for this snippet or you can start to correct this deadline because you know that's relevant for me and afterwards save it and also export it either in GES and extra documents which are still used because you want to work on these pages or to export PDF or whatever so that we can make something with you. So that's our really thinking. We'll see what we will really realize that probably is smiling because he knows that this will be unstable. I skip the demo and come to the sharing HDR models. Yes, I think you have seen that there are really a lot of models now available but the only ones who can see this are us and that's not a very satisfying situation. So I know that many of you want to share your models and that really makes sense and it also would, of course, lead to the situation that more people are working together with single users in Transcubus but that we are working together and that we get the feeling, oh wow, someone trained a great Latin model for medieval stuff and I am benefiting from that and that of course makes the group much more group and I think this is really great. So I'm happy that we are now in this situation that we can think about that and the question is of course how to make it. So I'll share some ideas on this. I think we will need some metadata so the language does not play a direct role for the HDR but on the other hand since the neural networks learn the context the distribution of characters will have an influence so language will be an idea then the kind of writing is actually very hard, often to say because it's just fussy medieval stuff or late 18th century but on the other hand we have heard today that people are training models for early writing, data writing and so on. Then the script, I mean definitely to print it and hand writing is something which makes sense but also specific writers styles maybe from the Middle Ages maybe in German, Korean, or Switzerland similar styles probably would exist in other languages as well. What is very helpful I think is the description of the document for training that's because the experts know what is specific with their document and other experts will understand exactly this thing, a short narrative for all the two lines I think will always be helpful but I think to have an example of the pages is of course the easiest way to share the knowledge models or to attract other users to use them but here we have to see that or we will see that of course some of you will be happy to share test data so if you have a test set you just say okay I'll share it to everyone then it's fine and we can show all the data to everyone but others will say well I got the images from a library I'm not sure what is the legal situation so I cannot say just actually act with everyone which means more or less like a publication or distribution so we came up with the idea maybe of sharing just a few lines which is likely enough so this interface now and here it would have more or less just a button that says share this model which is in the community and the idea is the lines could be that some random lines are selected from the document and if these lines do not contain any names or anything what you say is some more complications you will select some lines and the user will see at least some lines and add an idea of the one style of this model and the consideration is also that what this is really proved for us model that you just need to then scribe some lines and specify maybe the period of the language and then that there is a button to run this against all available public models and so that you get a base on the contours of several lines of the best model available for your specific document that I think can also be done with the most effort so these are some ideas on how we could do it and soon everyone of you will have the chance to have a look to this own document via the My Collections website so this means that you see all your documents on the web as well and then you can select the document and the document pages and here are the two main modes the plain text mode which means it's also mainly for transcription and it really allows you to concentrate on the transcription and you have the same options as with the possible expert type that you can add special characters that you can superscript the underline for and so on but you also can and this is also configurable or will be configurable that users can use annotation or an extended version of this tool and this means that the annotation can also be done in here including the properties so the attributes for the text and that's currently more or less all but we believe that with this we will be able to involve people who are not so much familiar or who do not want to get familiar with this expert type but want to complicate the infrastructure ok, thank you very much Günther it's the nerve of everyone in the crowd so let me just get a quick show of hands for open questions just a little question Günther do you think that it is possible that annotations are also part of the training that the machine itself in the future will recognize names etc so this problem is well known as name entity recognition and it depends on training data so we are of depth yeah, we really need good training data for name entity recognition or tagging and then we can maybe provide somewhere something I know that the company is working for us actually it's right now quite high at our priorities so we would be happy about some more we have some data sets some academic stuff but if there are a few more from the historical problems then we are we think we can provide something within the next half year thank you very much again Günther will be here as well as the other speakers