Spell Correction on Twitter dataset (test)
67.6Whole-Word AccuracySeq2Seq (Soft-T2T)
Evaluation results
| Method | Links | |
|---|---|---|
| Seq2Seq (Soft-T2T)Architecture=Transformer-based ED, Encoder layers=2, Decoder layers=22019.10 | 67.6 | |
| PIEArchitecture=4 layer self-attention transformer, Refinement iterations=1, Hidden units=2002019.10 | 67 | |
| Ribeiro2018 (best model)Model Type=local transduction model2019.10 | 54.1 | |
| Seq2Seq (Hard, lstm)Attention=hard monotonic, Architecture=LSTM2019.10 | 52.2 | |
| Seq2Seq (Soft, lstm)Attention=soft, Architecture=LSTM2019.10 | 46.3 |