Zilong Wang

Yes, we first sum up the "RD" for each page in the testing set and then divide it by the number of pages in the testing set so the final...

The LayoutReader uses the weight of LayoutLM(v1) as the initial states and is further trained on the ReadingBank for the reading order detection. Theoretically, it is possible to integrate the...

Hi! Thanks for your interest in our work! We are sorry to see that you couldn't produce the supposed results as we expected. All results reported in our paper are...

@SimeonZhang Hi, it is easy to run LayoutReader in `text only` settings by using the corresponding `model_name` and `model_name_or_path` in args. You can also run the `layout only` settings with...

@ManuelFay Thank you for the great job! As for your question, we run LayoutReader on ReadingBank dataset and we filter out the pages with more than 511 tokens. For other...

Actually, the [load_and_cache_line_order_examples](https://github.com/microsoft/unilm/blob/e4929f812398207b7fefb4dda6e9debcb8ce34b9/layoutreader/s2s_ft/utils.py#L339) is deprecated. You can reproduce a similar function if you need to conduct such experiments.

Our dataset, ReadingBank, is generated automatically using the xml meta data inside of the MS Doc files. It is described in detail in our paper (https://arxiv.org/pdf/2108.11591v2.pdf). Since there is no...

Thanks for your interest in our work! It is still an open question to do the doc AI in an end-to-end way. It must be a very promising direction and...

Hi, thanks for your interest in our paper. I'd love to help you to fix the issues. For the first question, I am not sure what the problem is without...

Hi @Mountchicken, the pre-trained model is based on the word-level. Unfortunately, I cannot help you with the textline-level pre-training.