This is the codebase for the paper Moving Down the Long Tail of Word Sense Disambiguation with Gloss Informed Bi-encoders.
Our bi-encoder model consists of two independent, transformer encoders: (1) a context encoder, which represents the target word (and its surrounding context) and (2) a gloss encoder, that embeds the definition text for each word sense. Each encoder is initalized with a pertrained model and optimized independently.
To run this code, you'll need the following libraries:
We used the WSD Evaluation Framework for training and evaluating our model.
To train a biencoder model, run python biencoder.py --data-path $path_to_wsd_data --ckpt $path_to_checkpoint. The required arguments are: --data-path, which is the filepath to the top-level directory of the WSD Evaluation Framework; and --ckpt, which is the filepath of the directory to which to save the trained model checkpoints and prediction files. The Scorer.java in the WSD Framework data files needs to be compiled, with the Scorer.class file in the original directory of the Scorer file.
It is recommended you train this model using the --multigpu flag to enable model parallel (note that this requires two available GPUs). More hyperparameter options are available as arguments; run python biencoder.py -h for all possible arguments.
To evaluate an existing biencoder, run python biencoder.py --data-path $path_to_wsd_data --ckpt $path_to_model_checkpoint --eval --split $wsd_eval_set. Without --split, this defaults to evaluating on the development set, semeval2007. The model weights and predictions for the biencoder reported in the paper can be found here.
Similar commands can be used to run the frozen probe for WSD (frozen_pretrained_encoder.py) and the finetuning a pretrained, single encoder classifier for WSD (finetune_pretrained_encoder.py).
If you use this work, please cite the corresponding paper:
@inproceedings{
blevins2020wsd,
title={Moving Down the Long Tail of Word Sense Disambiguation with Gloss Informed Bi-encoders},
author={Terra Blevins and Luke Zettlemoyer},
booktitle={Proceedings of the 58th Association for Computational Linguistics},
year={2020},
url={https://blvns.github.io/papers/acl2020.pdf}
}
Please address any questions or comments about this codebase to blvns@cs.washington.edu. If you want to suggest changes or improvements, please check out the CONTRIBUTING file for how to help out.
This codebase is Attribution-NonCommercial 4.0 International licensed, as found in the LICENSE file.