Table 4 BART model performance after layer manipulation.
From: Brain-model neural similarity reveals abstractive summarization performance
Ablated layer | Rouge1 | Rouge2 | RougeL | Perplexity | Noisy layer | Rouge1 | Rouge2 | RougeL | Perplexity |
|---|---|---|---|---|---|---|---|---|---|
w/o | 0.4570 | 0.2172 | 0.4839 | 6.5545 | w/o | 0.4570 | 0.2172 | 0.4839 | 6.5545 |
1 | 0.4536 | 0.2172 | 0.4213 | 6.9029 | 1 | 0.4571 | 0.2223 | 0.4249 | 6.5728 |
2 | 0.4476 | 0.2144 | 0.4159 | 7.4548 | 2 | 0.4557 | 0.2219 | 0.4239 | 6.6231 |
3 | 0.4497 | 0.2142 | 0.4171 | 7.7634 | 3 | 0.4545 | 0.2214 | 0.4224 | 6.6872 |
4 | 0.4514 | 0.2178 | 0.4194 | 6.8744 | 4 | 0.4577 | 0.2227 | 0.4256 | 6.5961 |
5 | 0.4443 | 0.2114 | 0.4127 | 7.3242 | 5 | 0.4551 | 0.2205 | 0.4227 | 6.7478 |
6 | 0.4349 | 0.2039 | 0.4041 | 7.6296 | 6 | 0.4507 | 0.2180 | 0.4188 | 6.9318 |
7 | 0.4417 | 0.2107 | 0.4102 | 7.2417 | 7 | 0.4540 | 0.2183 | 0.4218 | 6.8882 |
8 | 0.4301 | 0.2044 | 0.3993 | 7.5065 | 8 | 0.4501 | 0.2158 | 0.4182 | 6.9067 |
9 | 0.4356 | 0.2021 | 0.4039 | 7.8835 | 9 | 0.4507 | 0.2182 | 0.4192 | 6.7484 |
10 | 0.4409 | 0.2071 | 0.4090 | 7.2498 | 10 | 0.4517 | 0.2162 | 0.4207 | 6.3082 |
11 | 0.4337 | 0.1984 | 0.4029 | 6.5815 | 11 | 0.4425 | 0.2106 | 0.4110 | 6.4026 |
12 | 0.3840 | 0.1677 | 0.3572 | 6.9424 | 12 | 0.4491 | 0.2133 | 0.4174 | 5.9276 |
All | 0.0549 | 0.0024 | 0.0527 | 1.1475 | all | 0.2653 | 0.0763 | 0.2444 | 10.0152 |
