Transfer across rows
Our objective here is to analyze how training on one newpaper column for a particular category transfers to another column.
The column headers indicate which dataset BLEU scores are reported on, and the row labels indicate which column the model is trained on.
| bouncer train | bouncer test | bouncer dev | experts train | experts test | experts dev | ipl train | ipl test | ipl dev | match-report train | match-report test | match-report dev | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| bouncer | 87.44 | 52.50 | 52.70 | 45.18 | 46.11 | 45.13 | 51.98 | 52.95 | 52.67 | 54.38 | 54.60 | 54.43 |
| experts | 35.33 | 35.05 | 35.32 | 89.17 | 43.45 | 43.42 | 36.87 | 36.88 | 36.41 | 34.49 | 35.44 | 34.92 |
| ipl | 21.83 | 21.92 | 22.15 | 20.57 | 20.96 | 20.88 | 49.97 | 33.49 | 32.89 | 23.34 | 23.65 | 23.26 |
| match-report | 32.94 | 33.48 | 33.31 | 26.48 | 27.31 | 26.53 | 39.67 | 42.38 | 42.00 | 91.90 | 56.82 | 56.43 |
| all_en-hi | 73.94 | 54.47 | 54.74 | 75.11 | 52.72 | 52.31 | 78.02 | 63.36 | 62.7 | 80.76 | 64.29 | 64.1 |
| bouncer train | bouncer test | bouncer dev | experts train | experts test | experts dev | ipl train | ipl test | ipl dev | match-report train | match-report test | match-report dev | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| bouncer | 0.0 | 0.0 | 0.0 | -43.99 | 2.66 | 1.71 | 2.01 | 19.46 | 19.78 | -37.52 | -2.22 | -2.0 |
| experts | -52.11 | -17.45 | -17.38 | 0.0 | 0.0 | 0.0 | -13.1 | 3.39 | 3.52 | -57.41 | -21.38 | -21.51 |
| ipl | -65.61 | -30.58 | -30.55 | -68.6 | -22.49 | -22.54 | 0.0 | 0.0 | 0.0 | -68.56 | -33.17 | -33.17 |
| match-report | -54.5 | -19.02 | -19.39 | -62.69 | -16.14 | -16.89 | -10.3 | 8.89 | 9.11 | 0.0 | 0.0 | 0.0 |
| all_en-hi | -13.5 | 1.97 | 2.04 | -14.06 | 9.27 | 8.89 | 28.05 | 29.87 | 29.81 | -11.14 | 7.47 | 7.67 |