Together We make Sense–Learning of Meta-Sense Embeddings

This code in this repository is related to the ACL 2023 Findings Paper

Preparation

To reproduce our results, you may use the following command to create a conda environment.

conda update conda
conda env create -f environment.yml
conda activate npms

Word Sense Disambiguation(WSD)

Obtain meta embedding

• AVG

python3 base_method.py -i emb1 emb2 -o outfile -m avg

• CONC

python3 base_method.py -i emb1 emb2 -o outfile -m cat

• SVD

python3 base_method.py -m svd -k 2048 -i emb1 emb2 -o outfile

• AEME

The AEME is adopted from the open source for the COLING paper: Learning Word Meta-Embeddings by Autoencoding

git clone https://github.com/CongBao/AutoencodedMetaEmbedding.git

Follow the instruction in the code and use the following command to generate AEME meta embedding.

python run.py -m AAEME -i emb1 emb2 -d emb1_dim emb2_dim -o outfile --embed-dim 2048

• NPMS

First go to the neighbor directory.

cd neighbor

The alpha can be a fixed hyper-parameter specified using following command

python3  npms.py -i emb1 emb2  -alpha a 

To tune the value of alpha, we need to set the argument hyper to True

python3  npms.py -i -path emb1 emb2 -hyper True

Evaluate meta embedding

First switch to the eval_wsd directory

cd eval_wsd

For AEME and SVD, please run

python3 eval_proj.py -sv_path emb_path -test_set test_set_name -tran projection_matrix_path

For any other meta embedding methods

python3 eval.py -sv_path emb_path -test_set test_set_name

Word in Context(WiC)

This paper tackle WiC by training a classifier and give prediction using this model.

cd wic
python3 train_classifier.py -sv_path emb_path -out_path model_path -tran projection_matrix_path
python3 eval_classifier_wic.py -sv_path emb_path -clf_path model_path -tran projection_matrix_path