GraphWSD (TextGraphs-16 COLING 2022)

Official code repository: Word Sense Disambiguation of French Lexicographical Examples Using Lexical Networks

[Paper] [PPT]

Folder Description

/data/ : contains instructions to organize the data/ folder

/scripts/ : contains individual script modules

Steps to run the experiments:

1. Clone the repo : git clone https://github.com/ATILF-UMR7118/GraphWSD.git

2. Create the virtualenv

python3 -m venv wsdvenv
. wsdvenv/bin/activate
pip3 install --upgrade pip
cd GraphWSD/
pip3 install -r requirements.txt

3. Follow the instructions provided in data/ folder

4. To run the models for NOUN/VERB wsd:

(a.1) STRUCT model

python3  ~/GraphWSD/scripts/wsd_ewiser.py \
        --data ~/GraphWSD/data/ortolang/nountmp/ \
        --save_dir ~/GraphWSD/scripts/ortolog/ \
        --num 100  --model_num onoun_ewiser_29061156 --mtype ewiser --save-model \
        --learning 0.001  --hidden 8000 --batch 64 --device cuda --embed 768 --lm camembert-base

(a.2) STRUCT* model

python3  ~/GraphWSD/scripts/wsd_ewiser.py \
        --data ~/GraphWSD/data/ortolang/nountmp/ \
        --save_dir ~/GraphWSD/scripts/ortolog/ \
        --num 100  --model_num onoun_ewiser_29061156 --mtype ewiser --save-model \
        --learning 0.001  --hidden 8000 --batch 64 --device cuda --embed 768 --lm camembert-base --trainable

(a.3) STRUCT** model

python3  ~/GraphWSD/scripts/wsd_ewiser.py \
        --data ~/GraphWSD/data/ortolang/nountmp/ \
        --save_dir ~/GraphWSD/scripts/ortolog/ \
        --num 100  --model_num onoun_ewiser_29061156 --mtype ewiser --save-model \
        --learning 0.001  --hidden 8000 --batch 64 --device cuda --embed 768 --lm camembert-base --fragment --trainable

(b.1) SEM model

python3  ~/GraphWSD/scripts/wsd_ewiser.py \
        --data ~/GraphWSD/data/ortolang/nountmp/ --num 100 \
        --save_dir ~/GraphWSD/scripts/ortolog/  --model_num onoun_seml_29061522 \
        --mtype ewiserc --save-model --batch 64  --device cuda --semantics\
          --hidden-dim 8000  --embed 768 --lm camembert-base 

(b.2) SEM* model

python3  ~/GraphWSD/scripts/wsd_ewiser.py \
        --data ~/GraphWSD/data/ortolang/nountmp/ --num 100 \
        --save_dir ~/GraphWSD/scripts/ortolog/  --model_num onoun_seml_29061522 \
        --mtype ewiserc --save-model --batch 64  --device cuda --semantics\
          --hidden-dim 8000  --embed 768 --lm camembert-base --trainable

(b.3) SEM** model

python3  ~/GraphWSD/scripts/wsd_ewiser.py \
        --data ~/GraphWSD/data/ortolang/nountmp/ --num 100 \
        --save_dir ~/GraphWSD/scripts/ortolog/  --model_num onoun_seml_29061522 \
        --mtype ewiserc --save-model --batch 64  --device cuda --semantics\
          --hidden-dim 8000  --embed 768 --lm camembert-base --fragment --trainable

Different configurations

STRUCT and SEM are two strategies to intialize $A$ adjacency matrix. Any $a_{i,j} \in A$ is $\sum w(r)$ where $r \in S_{i,j}$ and $S_{i,j}$ is set of edges between i and j nodes.

STRUCT : count number of edges

SEM : count weight (strength) of edges. SEM model requires --semantics

config	interpretation	command
STRUCT/SEM	A is frozen
STRUCT/SEM *	$a_{i,j}$ is trainable	--trainable
STRUCT/SEM **	$w(r)$ is trainable	--trainable --fragment

For any questions related to repository contact: asinha@atilf.fr

Citation

@inproceedings{sinha2022word,
  title={Word sense disambiguation of french lexicographical examples using lexical networks},
  author={Sinha, Aman and Ollinger, Sandrine and Constant, Mathieu},
  booktitle={TextGraphs-16: Graph-based Methods for Natural Language Processing},
  pages={70--76},
  year={2022}
}