 Okay let's move on and we now have a presentation by Lyt Le Franc from the University of Antwerp which is also a member of the Reed Carp. All right here we go. Thank you. Good afternoon everyone my name is Lyt Le Franc and I'm working on my PhD in history at Antwerp University. My research is on the gendered experience of street life in nocturnal Antwerp and I use amongst others local police reports more specifically the incident books of Antwerp. Today I will talk to you about my attempts to train an HDR model for this historical source. First I will tell you a little bit more about the data set specifics. I've selected almost 300 incident books for HDR which dates from 1876 up to 1945. The incident books document reports by local police officers after patrolling 11 different districts in Antwerp. I've estimated that the fully digitized data set will contain about 30 and a half million words. The incident books are quite challenging for HDR for various reasons. Firstly they contain numerous different handwritings because the local police officers had to report individually on their rounds. Second they are written in French as well as Dutch and third they have a very complex layout which makes the segmentation very time consuming. The margins are for example often interpreted by transcribous as a separate text region which is problematic because the margins contain important information such as dates which is detached from the description of the incidents themselves and on the level of baselines transcribous often struggles with the recognition of skewed handwriting. This is my workflow which is pretty straightforward so I won't dwell on this. Perhaps interesting to note is the fact that I've experimented with the recent feature of transcribous to train baseline models and this enhanced the recognition of baselines already a little bit and I've also trained different models progressively so I added more training data to my models in order to be able to review their progress and on this slide you can see the first results of my training attempts for which I used SITLAB HDR plus and as a baseline model. In the table on top you can see the results of the models based on my own transcription work which I've called ground truth and I've composed my training sets by randomly selecting a few pages from each individual incident book and in order to be able to compare the different models on an equal basis I've composed a external validation set in the same manner. The low CR on the training sets is probably due to overfitting so I can't take this into account but the CR on the validation set is probably a more reliable parameter. I think it's quite promising that the results are improving by adding more training data but I'm a little bit disappointed that even though I already quadrupled my training set the CR on the validation set only dropped 1.8% up until now. In the previous academic year I have engaged some history students in my training endeavors and in the table below you can see the results based on their transcription work because it would have made no sense for their own research to transcribe random pages. I've selected two blocks of consecutive pages from two different sample periods. As I could have expected not all of their work is of the same quality. They struggled a lot with the segmentation and some of them were very lax with the tagging of unclear text so that's why I also decided to differentiate between the top four students, the top eight students and all of the students. The results were again less promising than I hoped. Despite the large work counts of the training sets they processed in transcribes and the lower amount of handwriting thanks to less sampling the student models don't perform below 10.63%. However surprisingly these models are better than the models I trained on the basis of my own transcription work. This observation could make me very insecure about my own transcription skills but it probably just proves that more training data so quantity is in fact more important than qualitative training data. So that's why I decided to combine the different training sets and I've experimented with different combinations and with different recognition engines which has led up to the best results so far. The lowest CR I've obtained was on the largest training sets with again SITLAB HDR+, using as a base model and it gave me the results of 9.29% on the validation sets. I'm still not satisfied with this result so that's why I try to identify the weaknesses of my model by analyzing the errors of the worst and the best models I've trained so far and this made me realize that both models actually make the same mistakes just in varying degrees. Unfortunately transcribes doesn't have a feature to systematically review the mistakes a model makes in detail so that's why I used an external tool to analyze the mistakes and that's how I learned that most mistakes were made with uppercase, spatial and special characters which with person names, street names and French words and with skew texts and texts in pencil. Now what's next? I'm still dreaming of a CR below 5% and I hope to obtain this by first speeding up the training process by developing a segmentation model for which I will probably use an external tool as well because the layout analysis transcribes provide unfortunately doesn't suffice for my sources because I also observed that more training data is still leading up to better results. I will expand my ground truth up to 150,000 words and I will also try to respond to weaknesses by experimenting with dictionaries, by excluding the skew texts. I will also switch to a more relaxed evaluation of my models, by excluding for example spatial characters or inter-function and in the last stage I will try to lower the CR even further by post-correcting my training data. If any of you has any other ideas or suggestions on how I can improve my models don't hesitate to contact me and thank you very much for your attention.