pretrained models using 163M data presented in "Various Errors Improve Neural Grammatical Error Correction"
-
bpe.model / bpe.vocab
- models of sentencepiece (0.1.95)
-
{1..5}.pt
-
data-bin/dict.{src,trg}.txt
- vocabulary of fairseq models
-
note
- Please normalize inputs before applying BPE using reguligilo (https://github.com/nymwa/reguligilo).
- Please use -a option like
reguligilo -a. - input -> reguligilo -> bpe -> encoder-decoder -> remove bpe -> malreguligilo (denormalization) -> output
-
scores (beam=12 lenpen=0.6)
| BEA19 | CoNLL 14 | JFLEG | |
|---|---|---|---|
| single 0 | 59.62 | 55.05 | 58.09 |
| single 1 | 59.71 | 55.29 | 58.01 |
| single 2 | 60.59 | 56.22 | 58.09 |
| single 3 | 59.76 | 55.35 | 58.56 |
| single 4 | 59.82 | 55.17 | 58.39 |
| average | 59.90 | 55.42 | 58.23 |
| ensemble | 60.89 | 55.73 | 58.37 |