language: - hr - bs - sr - cnr - hbs
BERTić [bert-ich] /bɜrtitʃ/ - A BERT model for Bosnian, Croatian, Montenegrin and Serbian
This Electra model was trained on more than 6 billion tokens of Bosnian, Croatian, Montenegrin and Serbian text.
Comparing this model to multilingual BERT and CroSloEngual BERT on the tasks of part-of-speech tagging, named entity recognition, geolocation prediction and choice of plausible alternatives shows this model to be superior to the other two.
Part-of-speech tagging
Evaluation metric is (seqeval) microF1. Reported are means of five runs. Best results are presented in bold. Statistical significance is calculated between two best-performing systems via a two-tailed t-test (* p<=0.05, ** p<=0.01, *** p<=0.001, ***** p<=0.0001).
| Dataset | Language | Variety | CLASSLA | mBERT | cseBERT | BERTić |
|---|---|---|---|---|---|---|
| hr500k | Croatian | standard | 93.87 | 94.60 | 95.74 | ***95.81 |
| reldi-hr | Croatian | internet non-standard | - | 88.87 | 91.63 | ***92.28 |
| SETimes.SR | Serbian | standard | 95.00 | 95.50 | 96.41 | 96.31 |
| reldi-sr | Serbian | internet non-standard | - | 91.26 | 93.54 | ***93.90 |
Named entity recognition
Evaluation metric is (seqeval) microF1. Reported are means of five runs. Best results are presented in bold. Statistical significance is calculated between two best-performing systems via a two-tailed t-test (* p<=0.05, ** p<=0.01, *** p<=0.001, ***** p<=0.0001).
| Dataset | Language | Variety | CLASSLA | mBERT | cseBERT | BERTić |
|---|---|---|---|---|---|---|
| hr500k | Croatian | standard | 80.13 | 85.67 | 88.98 | ****89.21 |
| reldi-hr | Croatian | internet non-standard | - | 76.06 | 81.38 | ****83.05 |
| SETimes.SR | Serbian | standard | 84.64 | 92.41 | 92.28 | 92.02 |
| reldi-sr | Serbian | internet non-standard | - | 81.29 | 82.76 | ****87.92 |
Geolocation prediction
Evaluation metrics are median and mean of distance between gold and predicted geolocations (lower is better). No statistical significance is computed due to large test set (39,723 instances). Centroid baseline predicts each text to be created in the centroid of the training dataset.
| System | Median | Mean |
|---|---|---|
| centroid | 107.10 | 145.72 |
| mBERT | 42.25 | 82.05 |
| cseBERT | 40.76 | 81.88 |
| BERTić | 37.96 | 79.30 |
Choice Of Plausible Alternatives (translation to Croatian)
Evaluation metric is accuracy. Best results are presented in bold. Statistical significance is calculated between two best-performing systems via a two-tailed t-test (* p<=0.05, ** p<=0.01, *** p<=0.001, ***** p<=0.0001).
| System | Accuracy |
|---|---|
| random | 50.00 |
| mBERT | 54.12 |
| cseBERT | 61.80 |
| BERTić | **65.76 |