Hi, thanks for feedback! I'll add more general information. In my opinion theres a lot to do in complex document classification, I'll try to add some demo to make things more intuitive. thanks!
the hardest part in training model in foreign languages is to get correctly labeled dataset. I worked with pretrain model on Polish language documents and based on this experience it is relatively good if you are using some text similarity measures. There are some examples/pretrain models with Korean/English/French language
In the example, text is not sorted by it's corrdinates but by appearence of boxes in first network. It is visible in more complex documents, that crnn network did not create boxes in descending order (word-by-word).
also, good point about the list. dictionary keys has no logical usage in this one.