Fig. 1.

The overall architectures of the proposed framework DTranNER. a As a CRF-based framework, DTranNER is comprised of two separate, underlying deep learning-based networks: Unary-Network and Pairwise-Network are arranged to yield agreed label sequences in the prediction stage. The underlying DL-based networks of DTranNER are trained via two separate CRFs: Unary-CRF and Pairwise-CRF. b The architecture of Unary-CRF. It is dedicated to train Unary-Network. c The architecture of Pairwise-CRF. It is also committed to train Pairwise-Network. A token embedding layer is shared by Unary-Network and Pairwise-Network. A token-embedding is built upon by concatenating its traditional word embedding (denoted as “W2V”) and its contextualized token embedding (denoted as “ELMo”)