| ================================================================================ |
| LANGUAGE IDENTIFICATION MODEL EVALUATION SUMMARY |
| ================================================================================ |
|
|
| Evaluation Date: 2026-01-31 15:03:05 |
| Model: Hybrid TF-IDF + BiLSTM Language Identifier |
| Dataset: WiLI-2018 |
| Number of languages: 235 |
| Vocabulary size: 20,002 |
| Total test samples: 117,500 |
|
|
| PERFORMANCE METRICS: |
| Test Accuracy: 93.7481% |
| Test F1 Score: 93.7531% |
|
|
| TOP 5 BEST PERFORMING LANGUAGES: |
| 1. ckb - 100.00% accuracy (500 samples) |
| 2. kbd - 100.00% accuracy (500 samples) |
| 3. min - 100.00% accuracy (500 samples) |
| 4. mlg - 100.00% accuracy (500 samples) |
| 5. bod - 99.80% accuracy (500 samples) |
|
|
| MOST CHALLENGING 5 LANGUAGES: |
| 1. wuu - 15.60% accuracy (500 samples) |
| 2. zh-yue - 22.80% accuracy (500 samples) |
| 3. zho - 37.00% accuracy (500 samples) |
| 4. hrv - 46.80% accuracy (500 samples) |
| 5. hbs - 53.80% accuracy (500 samples) |
|
|
| PERFORMANCE DISTRIBUTION: |
| Excellent (≥99%): 25 languages |
| Good (95-99%): 131 languages |
| Average (80-95%): 68 languages |
| Poor (<80%): 11 languages |
|
|
| INTERESTING FINDINGS: |
| 1. Several languages achieve 100% accuracy (ckb, kbd, min, mlg) |
| 2. Chinese variants are the most challenging (wuu: 15.6%, zh-yue: 22.8%) |
| 3. Japanese is surprisingly challenging (56.0% accuracy) |
| 4. 93.75% overall accuracy is excellent for 235 languages |
|
|
| RECOMMENDATIONS FOR IMPROVEMENT: |
| 1. Add data augmentation for low-accuracy languages |
| 2. Consider language family-based transfer learning |
| 3. Ensemble methods could boost performance |
|
|
| ================================================================================ |
|
|
