MDGI

The source code and data for ACL2021 Findings paper: Enhancing Metaphor Detection by Gloss-based Interpretations

The code is modified based on [Moving Down the Long Tail of Word Sense Disambiguation with Gloss Informed Bi-encoders] https://github.com/facebookresearch/wsd-biencoders

Dependencies

To run this code, you'll need the following libraries:

• Python 3
• Pytorch 1.2.0
• Pytorch Transformers 1.1.0
• Numpy 1.17.2
• NLTK 3.4.5
• tqdm
• xlrd 1.2.0

How to Run

To train a MDGI-Joint model, run python biencoder_PSUCMC_MDGI-Joint.py --data-path $path_to_data --ckpt $path_to_checkpoint --encoder-name $path_to_BERT(or other PLM). The required arguments are: --data-path, which is the access to the data path; and --ckpt, which is the filepath of the directory to which to save the trained model checkpoints and prediction files; --encoder-name, which is the name of the pretrained language model(if there exists problems, please look into the models/utils.py).

To evaluate an existing biencoder, run python biencoder_PSUCMC_MDGI-Joint.py --data-path $path_to_data --ckpt $path_to_checkpoint --eval .

Notice

1. MDGI-Joint-S is implemented by adding the argument --tie-encoders when running biencoder_*_MDGI-Joint.py.

2. When evaluating the model on VUA, please add the argument --context-max-length 256.

3. The current version does not support multigpu.