multilingual-WSD: A Python repository from suytingwan

This is the code repo for COLING2022: Multilingual Word Sense Disambiguation with Unified Sense Representation.

Envs

python 3.7 transformer 2.6

BabelNet, Inventory and Evaluation benchmark

BabelNet is a multilingual semantic lexicon which contains inventories for multiple languages. Inventories and multilingual WSD benchmark can be found at https://github.com/SapienzaNLP/mwsd-datasets. For MWSD evaluation, we use the wn split. After handling the inventories and evaluation dataset, the data are placed under ./data. The xml and gold key file in evaluation are further converted into csv format by:

cd ./preprocess
python convert_xml_csv.py

A small percentage (20%) is sampled from semeval-2013 evaluation dataset for model selection and placed under ./preprocess.

Machine Translation and Fastalign

First acquire the transalted corpora for languages DE, FR, IT and ES:

cd ./translation
python translate.py

Then, use fastalign tool to get the alignment between orginal semcor and translated corpora. This step will result alignment file under same folder.

Finally, generate training corpora for multiple languages by:

python mapping_synset.py

Online lexical knowledge gloss request

For experiments conducted on MuLaN dataset, the original data are filtered by inventories (WordNet split).

MuLaN contains some Babel synsets which do not exist in WordNet. For feasible experiments(collecting glosses according to inventories), we filter out the training instances labeled with words which are not shown in wn-split inventory generated from BabelNet.

cd mulan
python filter_mulan.py

For senses covered by annotated words but not shown in WordNet, we request BabelNet API to get gloss knowledge for the senses.

python request.py
python read_bn_gloss.py

Experiments

The difference between translated data and MuLaN is the different training data, inventories, and glosses. Since no validation data provided, a small percentage of test data from semevel13 is used as evaluation. For easy experiments, our code and preprocessed data can be download at code and data.

Experiments with translated data

CUDA_VISIBLE_DEVICES=3,0 python biencoder_mwsd.py --ckpt ./ckpts/trans_de \
     --data-path ./Evaluation \
     --valid-path ./preprocess \
     --train-path ./translation/train_corpora \
     --gloss-path ./mulan/mulan-de/gloss_new_combine.txt \
     --inventory-path ./mulan/mulan-de/de_inventory_filter_new_combine.txt \
     --grad-bsz 30 \
     --split semeval2013-de \
     --multigpu \
     --valid-small

Experiments with MuLaN

CUDA_VISIBLE_DEVICES=0,1 python biencoder_mwsd.py --ckpt ./ckpts/mulan_de \
     --data-path ./Evaluation \
     --valid-path ./preprocess \
     --train-path ./mulan/mulan-de/transfer_filter_new_check.csv \
     --gloss-path ./mulan/mulan-de/gloss_new_combine.txt \
     --inventory-path ./mulan/mulan-de/de_inventory_filter_new_combine.txt \
     --grad-bsz 30 \
     --split semeval2013-de \
     --multigpu \
     --valid-small

Evaluation

CUDA_VISIBLE_DEVICES=0,1 python biencoder_mwsd.py --ckpt ./ckpts/trans_de \
     --data-path ./Evaluation \
     --valid-path ./preprocess \
     --train-path ./translation/train_corpora \
     --gloss-path ./mulan/mulan-de/gloss_new_combine.txt \
     --inventory-path ./mulan/mulan-de/de_inventory_filter_new_combine.txt \
     --grad-bsz 30 \
     --split semeval2013-de \
     --multigpu \
     --eval

suytingwan/multilingual-WSD