rajicon / subatt

Implementation of SubAtt.

SubAtt

Implementation of SubAtt. For more information, see our paper : https://aclanthology.org/2023.findings-acl.221/

Quick Usage:

1. Build a corpus using preprocess from https://github.com/timoschick/form-context-model.

2. Call split_vocabulary_into_train_and_val.py on the preprocessed corpus folder to create a dict that splits training and validation words.

3. Train subword embeddings using train_torch_subwords.py

4. Train SubAtt model using train_SubAtt_All_Versions.py

5. Use trained model to predict new embeddings using its estimate_vector() and estimate_multiple_vectors() functions. Note: ANY input with zero subwords or contexts will currently lead to shape issues, so be sure to check for this before using these methods.

Some Examples:

python3 split_vocabulary_into_train_and_val.py $TrainCorpus #Step 2, creates TRAIN_VAL_DICT file

python3 train_torch_subwords.py $SubwordOutputDir $TrainCorpus $WordEmbs 30 0.7 1 64 1e-2 0 25 0 100 $TRAIN_VAL_DICT #Step 3

python3 train_SubAtt_All_Versions.py Pretrain_SubAtt $ModelOutputDir $TrainCorpus $SubwordOutputDir $WordEmbs 30 0.7 1 64 1e-4 0 25 0 100 8 fast 768 12 #Step 4

subatt's Issues

issue with Subword_Torch

Hello,

While following the instructions in the README.md to Step 3, I noticed that the "train_torch_subwords.py" script references a "Subword_Torch" model class, which does not seem to be defined in the codebase.

I suspect there might be a missing class definition or a typo in the class name.

Could you please give me some advice about how to resolve this issue?

Thank you for your time and patience.