提升BERT fine-tune正确率0.2%-0.3%的一个细节

字级别分词，不要用官方的tokenizer (https://github.com/google-research/bert/blob/master/tokenization.py)

自己重写一个

def tokenize_to_str_list(textString):
    split_tokens = []
    for i in range(len(textString)):
        split_tokens.append(textString[i])
    return split_tokens

def convert_to_int_list(split_tokens):
  output = []
  for token in split_tokens:
    if token in char2id: 
       output.append(char2id[item])
  return