BiLSTM together after softmax layer may be directly used to make a sequence labeling, but considering the legitimacy of the semantic context text conversion layer made CRF introducing certain constraints on the result of the network layer output BiLSTM to address similar to FIG It shows the problem:
The above figure label "I-Organization I-Person" this obvious error.
BiLSTM-CRF model two-layer structure, the first layer is a bidirectional LSTM layer, is responsible for automatic feature extraction sentence; the second layer is a layer of CRF, the tagging set of sentences, the use of dynamic decoding process in the Viterbi algorithm to find the optimal path.