Response Generation on DSTC7 Shared Task (test)
2.669NIST-4UNILM
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| UNILM2019.05 | 2.669 | 4.39 | 8.27 | 9.195 | 0.12 | 0.391 | 14.807 | — | — | |
| Human2021.10 | 2.65 | 3.13 | 8.31 | 10.44 | 16.65 | 67.01 | 18.73 | 2.63 | 12.36 | |
| Human Performance2019.05 | 2.65 | 3.13 | 8.31 | 10.445 | 0.167 | 0.67 | 18.76 | — | — | |
| Best System in DSTC7 Shared Task2019.05 | 2.523 | 1.83 | 8.07 | 9.03 | 0.109 | 0.325 | 15.133 | — | — | |
| Coherence Boostingalpha=-0.7, base_model=DialoGPT2021.10 | 1.79 | 0.94 | 5.55 | 9.78 | 28 | 72.46 | 16.63 | 1.78 | 6.33 | |
| DialoGPT (Greedy decoding)Inference method=Greedy decoding2021.10 | 1.63 | 1.72 | 6.78 | 6.45 | 6.19 | 17.56 | 13.3 | 1.62 | 9.92 | |
| Coherence Boostingalpha=-0.3, base_model=DialoGPT2021.10 | 0.73 | 3.53 | 6.91 | 8.54 | 16.81 | 49.35 | 9.75 | 0.72 | 13.82 | |
| DialoGPT (Beam search)Inference method=Beam search, beam_size=102021.10 | 0.02 | 3.23 | 5.35 | 6.06 | 14.03 | 34.59 | 5.81 | 0.02 | 12.81 |