Retrieval on MS MARCO
97.1RecallLabel Relaxation
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Label RelaxationModel Architecture=E5 model, Noise=0.12025.12 | 97.1 | — | — | — | |
| Label WeakeningModel Architecture=E5 model, Noise=0.02025.12 | 96.9 | — | — | — | |
| Label SmoothingModel Architecture=E5 model, Noise=0.02025.12 | 96.7 | — | — | — | |
| Label WeakeningModel Architecture=Dual Bi-Encoder Model, Noise=0.02025.12 | 89.6 | — | — | — | |
| Cross EntropyModel Architecture=Dual Bi-Encoder Model, Noise=0.02025.12 | 88.8 | — | — | — | |
| Label WeakeningModel Architecture=Dual Bi-Encoder Model, Noise=0.52025.12 | 81.2 | — | — | — | |
| LSNoise=0.0, Model=Cross-encoder, Loss=Label Smoothing2025.12 | 52.1 | — | — | 19.2 | |
| LWNoise=0.0, Model=Cross-encoder, Loss=Label Weakening2025.12 | 51.2 | — | — | 19.7 | |
| LRNoise=0.0, Model=Cross-encoder, Loss=Label Relaxation2025.12 | 50.9 | — | — | 19.6 | |
| AGCENoise=0.0, Model=Cross-encoder, Loss=Active Generalized Cross Entropy2025.12 | 50.6 | — | — | 19.3 | |
| NAGCENoise=0.0, Model=Cross-encoder, Loss=Combined NCE and AGCE2025.12 | 50.5 | — | — | 19.5 | |
| GCENoise=0.0, Model=Cross-encoder, Loss=Generalized Cross Entropy2025.12 | 50.4 | — | — | 19.1 | |
| AULNoise=0.0, Model=Cross-encoder, Loss=Adaptive Unhinged Estimator2025.12 | 50.4 | — | — | 19.9 | |
| NAULNoise=0.0, Model=Cross-encoder, Loss=Combined NCE and AUL2025.12 | 50.4 | — | — | 19.4 | |
| NCENoise=0.0, Model=Cross-encoder, Loss=Normalized Cross Entropy2025.12 | 50.1 | — | — | 18.8 | |
| ELSNoise=0.0, Model=Cross-encoder, Loss=Evidence-based Label Smoothing2025.12 | 50.1 | — | — | 19.4 | |
| CENoise=0.0, Model=Cross-encoder, Loss=Cross Entropy2025.12 | 49.9 | — | — | 18.6 | |
| LWNoise=0.5, Model=Cross-encoder, Loss=Label Weakening2025.12 | 46.9 | — | — | 16.9 | |
| CENoise=0.5, Model=Cross-encoder, Loss=Cross Entropy2025.12 | 45.1 | — | — | 14.8 | |
| Cross EntropyModel Architecture=Dual Bi-Encoder Model, Noise=0.52025.12 | 15.1 | — | — | — | |
| BFSevaluation_protocol=zero-shot2022.03 | — | 28 | 44.7 | — | |
| BM25evaluation_protocol=zero-shot2022.03 | — | 49 | 69 | — | |
| HLPtopology=DL (Dual-link), evaluation_protocol=zero-shot2022.03 | — | 42 | 62.6 | — | |
| HLPtopology=CM (Co-mention), evaluation_protocol=zero-shot2022.03 | — | 26.6 | 47.3 | — | |
| HLPtopology=DL+CM, evaluation_protocol=zero-shot2022.03 | — | 37.3 | 60 | — | |
| ICTevaluation_protocol=zero-shot2022.03 | — | 10.8 | 19.5 | — | |
| WLPevaluation_protocol=zero-shot2022.03 | — | 18.4 | 36 | — |