ighina / latin-ise-wsd

A large scale automatic analysis of selected lemmas sense change across centuries based on the Latin-ISE corpus and the original BERT-based word sense disambiguation system by Bamman et al. (2020)

Latin WSD experiments

This repository contains the code and the preliminary results of word sense disambiguation experiments in Latin.

The repository itself is a pseudo-fork of the original LatinBERT repository and it closely follows the word sense disambiguation setting of Bamman and Burns (2020), while using their BERT model pre-trained on latin corpora as described in the original paper.

Compared with Bamman and Burns (2020), we used all macro-senses from the Lewis and Short Latin-English dictionary. We ran the algorithm on the LatinISE corpus (McGillivray and Kilgarriff 2013).

B. McGillivray, A. Kilgarriff, Tools for historical corpus research, and a corpus of Latin, in: P. Bennett, M. Durrell, S. Scheible, R. J. Whitt (Eds.), New Methods in Historical Corpus Linguistics, Narr, Tübingen, 2013.

Install

Tested on Python 3.8.12 and 3.7.12.

1.) Create a conda environment (optional):

conda create --name latinbert python=3
conda activate latinbert

2.) Install PyTorch according to your own system requirements (GPU vs. CPU, CUDA version): https://pytorch.org.

3.) Install the remaining libraries:

pip install -r requirements.txt

4.) Install Latin tokenizer models:

python3 -c "from cltk.data.fetch import FetchCorpus; corpus_downloader = FetchCorpus(language='lat');corpus_downloader.import_corpus('lat_models_cltk')"

Use

All the code related to the word sense disambiguation analysis is contained in the wsd folder of this same repository, while this parent directory contains the pre-trained LatinBERT and other utils functions. See https://github.com/Ighina/Latin-ISE-WSD/tree/master/wsd#readme for details.