notesum.ai
Published at December 5A Context-aware Framework for Translation-mediated Conversations
cs.CL
Released Date: December 5, 2024
Authors: José Pombal1, Sweta Agrawal2, Patrick Fernandes3, Emmanouil Zaranis3, André F. T. Martins4
Aff.: 1Unbabel; 2Instituto de Telecomunicações Instituto Superior Técnico, Universidade de Lisboa; 3Instituto de Telecomunicações; 4University of Carnegie Mellon

| en-xx | xx-en | |||||||
| Model | Context? | chrF | Comet | MetricX | chrF | Comet | MetricX | |
| Baselines | ||||||||
| GPT-4o | ✗ | 70.09 | 92.62 | 0.37 | 77.33 | 92.41 | 0.50 | |
| ✓ | 70.34 | 92.93 | 0.33 | 74.75 | 91.59 | 0.58 | ||
| \cdashline1-9[.4pt/2pt] TowerInstruct | ✗ | 64.95 | 91.69 | 0.38 | 76.04 | 92.17 | 0.56 | |
| ✓ | 63.39 | 91.09 | 0.49 | 74.32 | 91.36 | 0.60 | ||
| + QAD (Comet) | ✗ | 65.20 | 92.87 | 0.31 | 75.59 | 92.80 | 0.52 | |
| + QAD (ContextComet) | ✗ | 65.06 | 92.57 | 0.31 | 75.91 | 92.65 | 0.51 | |
| \cdashline1-9[.4pt/2pt] TowerChat | ✗ | 71.68 | 93.01 | 0.32 | 77.97 | 92.72 | 0.51 | |
| ✓ | 75.93 | 93.63 | 0.32 | 78.87 | 93.01 | 0.47 | ||
| + QAD (Comet) | ✓ | 76.36 | 94.18 | 0.25 | 78.92 | 93.39 | 0.44 | |
| + QAD (ContextComet) | ✓ | 76.56 | 94.05 | 0.26 | 78.92 | 93.24 | 0.44 | |