Consistency between prediction and VALID_HTML_TOKEN

[ygzhu]

Thanks for your contributions! I noticed that the VALID_HTML_TOKEN is composed of [""] and HTML_TOKENS in constant.py. VALID_HTML_TOKEN is used to filter the prediction during inference phase. But the annotation of "[" and "]" are not included in VALID_HTML_TOKEN. Does this omission affect the final prediction during the inference phase?