Natural Language Generation on WoW (Human Evaluation)
4.68Mean RelevanceBM
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| BMDecoding method=Beam search, Beam size=52022.04 | 4.68 | 0.12 | 4.2 | 0 | 4.48 | 0.18 | |
| GoldDescription=Human reference2022.04 | 4.62 | — | 4.72 | — | 4.53 | — | |
| KIDDecoding method=Knowledge Infused Decoding2022.04 | 4.57 | 0.3 | 4.53 | 0.29 | 4.33 | 0.08 | |
| SPDecoding method=Sampling, top-p=0.9, k=202022.04 | 4.44 | 0.1 | 4.38 | 0.06 | 4.4 | 0.1 | |
| RFLCDecoding method=Reflective decoding2022.04 | 4.35 | 0.06 | 4.41 | 0.15 | 4.14 | 0.05 |