search

Hyperparticle / udify

A single model that parses Universal Dependencies across 75 languages. Given a sentence, jointly predicts part-of-speech tags, morphology tags, lemmas, and dependency trees.
https://arxiv.org/abs/1904.02099
MIT License
219 stars 56 forks source link

Frozen features and gold tags #2

Closed niless closed 4 years ago

niless commented 4 years ago

Hi,

Is it possible to disable fine-tuning by simply changing requires_grad to false in token_embedder for a frozen feature based embedder? Also, is there a preferred approach to just evaluate dependency parsing using gold tags for pos/feats/lemmas?

Thanks.

Hyperparticle commented 4 years ago

Sure, if you want to prevent your feature embedder from updating its weights, set requires_grad to false before passing it in.

What exactly are you evaluating? Are you saying you want to evaluate a predicted dependency tree just from the post/feats/lemma gold data alone? Seems like you'll need the gold dependency tree if you want a good evaluation, as it's globally context sensitive, while the other annotations don't need as much context.