search

google / sling

SLING - A natural language frame semantics parser
Apache License 2.0
1.93k stars 268 forks source link

Needed guideline: Using emb-vectors on input #265

Closed gopa810 closed 6 years ago

gopa810 commented 6 years ago

I would need to omit using word IDs, suffix IDs and word features and insert embedding vector directly to the input of LSTMs. Can you guide me where to change and what? Should I use FixedFeature or LinkedFeature for this purpose?

ringgaard commented 6 years ago

Which version of SLING are you using? The SEMPAR version in the main branch, or the new CASPAR version in the caspar branch?

gopa810 commented 6 years ago

Which version of SLING are you using? The SEMPAR version in the main branch, or the new CASPAR version in the caspar branch?

SEMPAR

ringgaard commented 6 years ago

SEMPAR is using the Dragnn framework, and it only support discrete features out of the box. However, if your embedding vectors are word-based, you could use these to initialize the pretrained embeddings and disable the other lexical features.

We are about to deprecate SEMPAR at the moment, and make CASPAR the main branch. CASPAR is using PyTorch, so it is easier to modify the code to support input using embedding vectors. Basically you will get rid of the features part of the network and input the embedding vectors directly into the LSTM.

gopa810 commented 6 years ago

Thanks.