Research Article
Arabic Syntactic Diacritics Restoration Using BERT Models
Table 5
The comparison between the proposed BERT model and the state-of-the-art systems on the ATB test dataset.
| | Method | All WER | Morph WER | CEER (%) |
| | MaxEnt tagger [10] | 18.00% | 7.90% | 10.10 | | Rule-based tagger [41] | — | — | 9.97 | | MADA tagger [42] | 14.90% | 5.50% | 9.40 | | Random forest tagger [28] | 13.70% | 4.30% | 9.40 | | Scoring of a language model [5] | 12.50% | 3.10% | 9.11 | | Confused subset resolution [43] | 11.60% | 3.00% | 8.60 | | Scoring of a language model [16] | 10.87% | 3.00% | 7.87 | | SVM tagger [29] | — | — | 6.8 | | MADAMIRA + character RNN tagger [21] | 8.40% | 2.30% | 6.10 | | Character RNN tagger [20] | 9.07% | 4.34% | 4.73 | | Word level MaxEnt/BiLSTM tagger [29] | — | — | 5.3 | | Word level MaxEnt/BiLSTM tagger + distillation of knowledge + embeddings based on characters [4] | — | — | 4.3 | | BERT tagger (two steps fine-tuning) | — | — | 2.94 |
|
|
