Bert in text representation effect is very good, why many NLP paper also combine Bert other model? Such as Bert + BiGRU etc, what is the function of BiGRU here, is to deep Bert said after the text feature extraction?