Table 2 Averaged results with standard deviations on four domains, where \(d_1 \rightarrow d_2\) indicates the domain adaptation from \(d_1\) (domain of \(D_d\)) to \(d_2\) (domain of \(D_t\)).
From: Subset selection for domain adaptive pre-training of language model
Domain adaptation | Task dataset | Vanilla model (w/o further pre-training) | RandomSet | SuzukiSet | AlignSet | FullSet |
|---|---|---|---|---|---|---|
BioMed \(\rightarrow\) BioMed | RCT | 85.71 ± 0.06 | 86.47 ± 0.08 | 86.39 ± 0.08 | 86.52 ± 0.04 | - |
CS \(\rightarrow\) CS | ACL-ARC | 74.07 ± 5.60 | 75.71 ± 2.01 | 75.56 ± 2.48 | 76.15 ± 1.51 | - |
News \(\rightarrow\) News | AGNEWS | 93.73 ± 0.19 | 93.90 ± 0.20 | 93.99 ± 0.06 | 94.12 ± 0.13 | 93.84 ± 0.13 |
HYPERPARTISAN | 81.81 ± 4.67 | 84.84 ± 4.35 | 84.95 ± 1.26 | 86.45 ± 1.98 | 82.63 ± 2.10 | |
Personality \(\rightarrow\) Personality | First Impressions V2 | 0.1128 ± 0.0013 | 0.1120 ± 0.0004 | 0.1121 ± 0.0004 | 0.1118 ± 0.0001 | – |
\(\dagger\) Personality \(\rightarrow\) Review | HELPFULNESS | 68.40 ± 0.65 | 68.21 ± 0.39 | 68.19 ± 0.28 | 68.42 ± 0.46 | – |
IMDB | 93.64 ± 0.01 | 93.32 ± 0.20 | 90.95 ± 0.24 | 93.35 ± 0.04 | – |
