Table 2.

Performance comparison of Semi-Supervised bi-LSTM (SS-BLSTM) under different word embedding initialization settings and different unlabeled data settings. Results are reported averaged over 30 trials along with the std. deviation

Method	F1-Score	Precision	Recall
SS-BLSTM (with drug mask removed)	0.747 ±0.037	0.723 ±0.106	0.780 ±0.108
SS-BLSTM (with labeled tweets dictionary only)	0.745 ±0.039	0.727 ±0.072	0.769 ±0.097
SS-BLSTM (with GoogleNews [25] vectors)	0.736 ±0.031	0.708 ±0.095	0.774 ±0.118
SS-BLSTM (with medical embeddings)	0.673 ±0.021	0.642 ±0.089	0.716 ±0.118

Highlighted portions reflect the best results across the respective columns