chunked/nltk_chunking/model_doc_visual_bert/chunk_21.txt

The text input is concatenated in the front of the visual embeddings in the embedding
layer, and is expected to be bound by [CLS] and a [SEP] tokens, as in BERT.