Does the method which called get_bert_embed_matrix that return `word_embed_6 = torch.cat([hidden_states[i] for i in [-1,-2,-3,-4]], dim=-1)` for sub-word?