Relation extraction results on the TACRED test partition. We used the pre-trained SpanBERT-large. Our full model trains on the entire training partition using the SSL method discussed in Section 3.2.1. The “burn-in only” setting trains just on the training subset that has annotations from rules.
| Approach . | Precision . | Recall . | F1 . |
|---|---|---|---|
| Baselines | |||
| Rules | 85.82 | 24.21 | 37.77 |
| SpanBERT (Joshi et al. 2020) | 69.97 ± 0.58 | 70.20 ± 1.73 | 70.07 ± 0.73 |
| Unsupervised Rationale | 69.24 ± 0.40 | 69.05 ± 1.86 | 69.14 ± 0.83 |
| Our Approach | |||
| Burn-in Only | 51.06 ± 3.57 | 48.32 ± 2.33 | 49.61 ± 2.42 |
| Full Model | 72.02 ± 0.90 | 69.11 ± 1.82 | 70.52 ± 0.54 |
| Approach . | Precision . | Recall . | F1 . |
|---|---|---|---|
| Baselines | |||
| Rules | 85.82 | 24.21 | 37.77 |
| SpanBERT (Joshi et al. 2020) | 69.97 ± 0.58 | 70.20 ± 1.73 | 70.07 ± 0.73 |
| Unsupervised Rationale | 69.24 ± 0.40 | 69.05 ± 1.86 | 69.14 ± 0.83 |
| Our Approach | |||
| Burn-in Only | 51.06 ± 3.57 | 48.32 ± 2.33 | 49.61 ± 2.42 |
| Full Model | 72.02 ± 0.90 | 69.11 ± 1.82 | 70.52 ± 0.54 |