Fig. 2
From: A hybrid deep learning framework for bacterial named entity recognition with domain features
The method to get char-level embedding in our paper. The characters in a word are transfer to vectors, then though a convolution layer and a max-pooling, finally the output are concatenated to represent the word