Semantic Textual Similarity on KorSTS (Perturbation Robustness)
77.42STS Score (Clean)KOMBO
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| KOMBOTokenization=Jamo2026.04 | 77.42 | 73.13 | 69.19 | 52.34 | 31.64 | 73.21 | 68.17 | 52.47 | 32.42 | |
| BERTTokenization=MorSubword2026.04 | 77.29 | 66.09 | 57.26 | 57.21 | 34.89 | 68.04 | 57.42 | 57.09 | 36.29 | |
| BERTTokenization=Subword2026.04 | 76.41 | 64.26 | 57 | 50.52 | 31.46 | 68.43 | 62.64 | 49.94 | 32.22 | |
| BERTTokenization=Character2026.04 | 76.34 | 70.93 | 66.33 | 49.86 | 32.3 | 72.34 | 68.95 | 49.63 | 33.81 | |
| BERTTokenization=Jamo2026.04 | 72.96 | 71.67 | 69.61 | 47.68 | 29.77 | 72.15 | 69.62 | 47.31 | 30.84 |