the d2e model has higer score than d2s model. is it normal ?

[wangwang110]

thank you for your good work. In my experiment(traing stage), I find a large gap between d2e and d2s and the d2e has higer score. is it normal ?

d2e perfomance:

{
"epoch": 50,
"scores": {
"train/loss": {
"loss": 50.01665752796086
},
"valid/score": {
"span": 0.8564028180687941,
"ns": 0.7262743472855366,
"relation": 0.5683796104434314,
"full": 0.5623704931620389
}
}
}

d2s performance:
{
"epoch": 50,
"scores": {
"train/loss": {
"loss": 29.35438394074393
},
"valid/score": {
"span": 0.6909492273730684,
"ns": 0.4586092715231788,
"relation": 0.2886313465783664,
"full": 0.2814569536423841
}
}
}

[KobayashiNaoki]

Yes, it is because training sets of D2S and D2E are different.
D2S is too small compared with D2E.
It's difficult to obtain high scores by D2S due to data sets size.

If you can increase or augment the data size, it can be useful for D2S settings.
(We don't use external data sets in our experiments. If you can use it, it would be better, I think. )

[wangwang110]

Yes, it is because training sets of D2S and D2E are different.
D2S is too small compared with D2E.
It's difficult to obtain high scores by D2S due to data sets size.

If you can increase or augment the data size, it can be useful for D2S settings.
(We don't use external data sets in our experiments. If you can use it, it would be better, I think. )

thank you for your reply, I understand.
but I also parser some raw texts, I find it seems all text got same structure at sentence level (not EDU), is it normal?

[YTZ01]

Hello, which package did you use for visualization? Is that nltk? @wangwang110