Table 1:
Prediction results of the different architectures on relation extraction (i.e., predicting whether the relation is R1 or R2), where P(Rk), R(Rk), F(Rk) are precision, recall and f-score and R1 = “not_located_in”, R2 = “located_in”. Similar values were obtained by conducting several experiments and averaging the results.
| Experiment | P(R1) | R(R1) | F(R1) | P(R2) | R(R2) | F(R2) |
| Word2Vec-CNN | 0.77 | 0.84 | 0.81 | 0.82 | 0.74 | 0.77 |
| Word2Vec-Bi-GRU | 0.90 | 0.88 | 0.89 | 0.77 | 0.80 | 0.78 |
| ELMo-Bi-LSTM | 0.92 | 0.90 | 0.91 | 0.80 | 0.85 | 0.83 |