2. MODELS TO COMPARE WE CONDUCTED EXPERIMENTS TO COMPARE THE PERFORM...

5.2. Models to compare

We conducted experiments to compare the performance of the models presented in

Table 6 using the method described in Section 5.1. The baseline model uses CRFs

with manually designed features. Our purpose is to investigate the task by using a

traditional statistical learning model

Table 6. Models to compare Model Word layer Sentence layer Inference layer Baseline CRFs CNNs-BiLSTM-Softmax CNNs BiLSTM Softmax CNNs-BiLSTM-CRFs CNNs BiLSTM CRFs BiLSTM-BiLSTM-Softmax BiLSTM BiLSTM Softmax BiLSTM-BiLSTM-CRFs BiLSTM BiLSTM CRFs

Note that for each of neural models (CNNs-BiLSTM-Softmax, CNNs-BiLSTM-

CRFs, BiLSTM-BiLSTM-Softmax, BiLSTM-BiLSTM-CRFs), we conducted

experiments with two variants of the model: 1) using only automatically learned

features; 2) using both automatically learned and manually designed features. The

purpose is to investigate the impact of manually designed features on the performance

of neural models.