how can we remove embedding layer ...

[SeekPoint]

for example, we want training the character model in some Asia language like Chinese..

[ajfisch]

You don't need to remove the embedding layer to train a character model. You can simply switch your tokenizer in preprocessing to split by character. Kim et al. still find that character embeddings (rather than sparse coding) is helpful (https://arxiv.org/abs/1508.06615).

The bottom line is, however, that if you want to change the architecture of the DocReader, you can always subclass it, use a different network architecture than defined in the rnn_reader.py, or provide an entirely new class to the pipeline, etc.

[SeekPoint]

"simply switch your tokenizer in preprocessing to split by character " is easy, but should I provide the embedding file like a "Chinese version" glove.840b.300d, and how to make it?

[ajfisch]

fasttext.cc has Chinese embeddings. We do not provide Chinese training corpora here though. DrQA only supports English.