This repository contains code and pretrained baselines (coming soon) for the xl-wic task released in this paper. Results are slightly different from what was reported on the paper. Please consider these figures when comparing on XL-WiC dataset.
| BG | DA | ET | FA | HR | JA | KO | NL | ZH | IT | FR | DE | AVG | Epoch | Batch Size | EN Dev | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| XLM-R Large | 62.05 | 66.27 | 66.15 | 78.25 | 64.71 | 57.28 | 69.92 | 70.22 | 61.20 | 58.28 | 60.20 | 61.31 | 64.65 | 8 | 32 | 74.14 |
The model averages subword embeddings to create word embeddings and concatenate the embeddings of the two target words before prediction (as also reported in the paper).