[QUESTION] Assigning weights to sentences in depparse training

[soras]

Question about training stanza depparse (on tokenized and pretagged document): is there a way to assign weights to input sentences, so that some sentences will appear more important than others? A naive approach would be to duplicate sentences, but I wonder if there is another way to do it?

[AngledLuffa]

It would be possible to add that feature. The loss gets calculated by iterating over the sentences, using the model to calculate the loss when the gold values are available:

https://github.com/stanfordnlp/stanza/blob/a85cce6816f40fa03ab06f6497c7d65ba1244a33/stanza/models/parser.py#L209 https://github.com/stanfordnlp/stanza/blob/a85cce6816f40fa03ab06f6497c7d65ba1244a33/stanza/models/depparse/trainer.py#L53

The model calculates individual pieces of the loss here:

https://github.com/stanfordnlp/stanza/blob/a85cce6816f40fa03ab06f6497c7d65ba1244a33/stanza/models/depparse/model.py#L190 https://github.com/stanfordnlp/stanza/blob/a85cce6816f40fa03ab06f6497c7d65ba1244a33/stanza/models/depparse/model.py#L196 https://github.com/stanfordnlp/stanza/blob/a85cce6816f40fa03ab06f6497c7d65ba1244a33/stanza/models/depparse/model.py#L209

You could weight it by whatever you like, as long as you pass in the sentence weights to model.forward(). I would say we're not going to add that feature any time soon, but if you're inspired to make a PR which adds that feature, we can talk about what it would look like

[soras]

Thank you for the quick and through reply! Unfortunately, adding this feature is out of my scope for the time being, so I am closing this issue now.

[AngledLuffa]

That's fair. We can leave this open, and hopefully one day one of us (or you) can add it as a feature.