Word Sense Extension

This is the GitHub repository containing code implementations and data for the following ACL 2023 paper:

Yu, L. and Xu, Y. (2023) Word sense extension. In Proceedings of the The 61st Annual Meeting of the Association for Computational Linguistics.

Data

See the 8 .csv files under the /data/ directory for word usage data used for WSE and WSD experiments. These dataframes can be generated by running the code under the /code/data_prep/ directory.

Code for data preparations

Python files for preparing the usage data are located under the /code/data_prep/ directory:

-- preprocess.py: code for preprocessing the wikitext-103 corpus using SpaCy.

-- wsd.py: code for annotating word sense labels the preprocessed wikitext-103 corpus through automated WSD models.

-- sense_usage_data_prep.py: code for preparing usage dataframes for WSE and WSD experiments.

-- cse_data_prep.py: code for preparing datasets for sense extensional space learning.

Code for word sense extension (WSE) experiments

Python files for WSE experiments are located under the /code/wse/ directory:

-- pretrain_mlm.py: code for pretraining BERT-based language models from scratch on usage data of partitioned tokens.

-- wse_train.py: code for learning sense-extensional semantic space and evaluating the learned WSE models.

Code for word sense disambiguation (WSD) experiments

Python files for WSD experiments are located under the /code/wsd/ directory:

-- wsd_models.py: implementations of the original BEM WSD model in (Blevins et al., 2020).

-- wsd_main.py: code for training and evaluating BEMs with and without sense-extensional space learning on the standard WSD evaluation framework by (Raganato et al., 2017).