 So what we did actually, so first, what is wujud? Wujud is a large corpus, about half million tokens, when annotated with misted name entities, misted. So when we say Edouard Said Foundation, Edouard Said Foundation is an organization and Edouard Said which is a sub-text is considered person, so that's wujud. So we went another living, basically, and we did it with finding more. So again, again, this is, you can scan this, in all my presentations, I showed this to show you where it's in the resources, where you can find it, and you can try download all open source. So extended wujud, we call it wujud fine. So wujud fine is an extension to wujud. So what we did, we say a sub-text. So I don't want to say, for example, Birzat University is an organization. No, I need to know what type of organization it is. So this is some examples. Nayami is a capital of Nigeria. So in wujud, we say this is GBE, GBE means geopolitical entity. Okay? And also to align with the Algerians' priorities in the region. So we also call it GBE. So now we came to say, okay, but this is a country, Nigeria is a country. So we want to know not only that it is geopolitical entity, we want to know that it's a country, and that this is actually, it's also GBE, but it's an organization. Here we are talking Nigeria, the government, not Nigeria, the country. So this is example. Another example is a picture of some patients at Al-Shifa Hospital. So it's an organization. Al-Shifa Hospital is an organization, or you can call it also a facility, but here, so what we did, we say it is the facility of the organization, basically. Another example, I live northeast of Gaza City. So this is, please notice that Gaza City itself is a GBE. Northwest of Gaza City is a location together. So now with the wujud fine, what we did, we say, okay, that one is town. This one is a region general, and so on. Another, this is another example. Birzette University announces scholarships for PhD students. So Birzette, by the way, Birzette is a town in Palestine, if you don't know, very nice town. So Birzette is a GBE, but it's also a town. Birzette University is an organization that is also an education. Okay, I will come to this, but so this is the tree we did. We took only four types of name entities, and we classify them into more types of entities. By the way, these entities are from the ACE Corpus, developed by LDC, but we enhanced it. We think we made some change. So coming back, so before annotating with subtypes, we had to change the wujud guidelines because to make it combatable. By the way, just to open the bracket, forget my presentation. I can see a research problem, a nice research problem in the name entity recognition area, which is compatibility of guidelines. This is a data engineering problem, not a natural language processing problem, but it should be really solved because all guidelines are not combatable, but that's another thing. Now, so these are statistics about the number of entities we revised before annotating with subtypes. Okay, these are the guidelines of the subtypes. You can look at the paper, and then we trend a model. We trend actually three, first week. I wasn't planning to be early, but it's okay. So we used different retrend models, but we first trend on flat without subtypes, without subtypes. Then we trend with misted, without subtypes. Then we trade misted and subtypes. And as you see here, so it's 88% accuracy, can I close the mouse? Oh, okay, 88.5% accuracy or if one score what we did. We did something actually a little more. We collected three domains out of, so we did an out-of-domain experiment, not seen at all this data. We collected just a small dataset for testing. We collected from Al Jazeera about bulletics and another science and finance. And you see the model actually, this is a limitation. The model is really has low numbers, especially in science domain. And I think this is a problem in most of the, all of this deep learning models, just to show it. Again, you can try download and see all subtypes and everything from this link. And so in GitHub and hugging face, all models, everything we have is open. Thank you. Time for one question, please. Okay, just quickly. The CJK Dictionary Institute, we specialize in Chinese and Japanese entity recognition. And I wanna ask a question. What we found the most difficult of all was what's called POI, Points of Interest, not towns, but things like the names of squares, the names of hotels, the kind of things that a tourist would find on a map. We found it very difficult to identify them and to translate them into other languages. Have you had any issues with this with POI, do you? Is that one of your categories that has been included in your research? So I will generalize you the question if you allow me to say, what's the difference between entity recognition in English and in Arabic in this deep learning era? It's actually, I think in Arabic is more challenge. Because it's difficult because there is no capitalization, names can be adjectives and so on. I don't know in Chinese, I cannot answer you, but in Arabic and in English, so it's language is missing, I would say. That's all I can say. I cannot answer you regarding Chinese. All right, I think we should move to the next because we're really...