🚀 Feature request

Fast Tokenizer for DeBERTA-V3 and mDeBERTa-V3

Motivation

DeBERTa V3 is an improved version of DeBERTa. With the V3 version, the authors also released a multilingual model "mDeBERTa-base" that outperforms XLM-R-base. However, DeBERTa V3 currently lacks a FastTokenizer implementation which makes it impossible to use with some of the example scripts (They require a Fa

From paper, it mentioned

Instead, the training data generator chooses 15% of tokens at random, e.g., in the sentence my
dog is hairy it chooses hairy.

It means that 15% of token will be choose for sure.

From https://github.com/codertimo/BERT-pytorch/blob/master/bert_pytorch/dataset/dataset.py#L68,
for every single token, it has 15% of chance that go though the followup procedure.

_handle_duplicate_documents and _drop_duplicate_documents in the elastic search document store will always report self.index as the index with the conflict, which is obviously incorrect.

Edit: Upon further investigation, this is actually a lot worse. Using multiple indices with the ElasticSearch DocumentStore is completely broken due to the fact, that this is used in `_handle_duplicate_do

目前的多音字使用 pypinyin 或者 g2pM，精度有限，想做一个基于 BERT (或者 ERNIE) 多音字预测模型，简单来说就是假设某语言有 100 个多音字，每个多音字最多有 3 个发音，那么可以在 BERT 后面接 100 个 3 分类器（简单的 fc 层即可），在预测时，找到对应的分类器进行分类即可。
参考论文：
tencent_polyphone.pdf

数据可以用 https://github.com/kakaobrain/g2pM 提供的数据

进阶：多任务的 BERT
![image](https://user-images.githubusercontent.com/24568452

Issue to track tutorial requests:

• Deep Learning with PyTorch: A 60 Minute Blitz - #69
• Sentence Classification - #79