F-Coref: Fast, Accurate and Easy to Use Coreference Resolution
F-Coref allows to process 2.8K OntoNotes documents in 25 seconds on a V100 GPU (compared to 6 minutes for the LingMess model, and to 12 minutes of the popular AllenNLP coreference model) with only a modest drop in accuracy. The fast speed is achieved through a combination of distillation of a compact model from the LingMess model, and an efficient batching implementation using a technique we call leftover
Please check the official repository for more details and updates.
Experiments
| Model | Runtime | Memory |
|---|---|---|
| Joshi et al. (2020) | 12:06 | 27.4 |
| Otmazgin et al. (2022) | 06:43 | 4.6 |
| + Batching | 06:00 | 6.6 |
| Kirstain et al. (2021) | 04:37 | 4.4 |
| Dobrovolskii (2021) | 03:49 | 3.5 |
| F-Coref | 00:45 | 3.3 |
| + Batching | 00:35 | 4.5 |
| + Leftovers batching | 00:25 | 4.0 |
| The inference time(Min:Sec) and memory(GiB) for each model on 2.8K documents. Average of 3 runs. Hardware, NVIDIA Tesla V100 SXM2. |