/WinoGrande

Coreference resolution with RoBERTa and GCN's

Primary LanguagePython

"Winobert"

Trained a model which employs a RoBERTa with token dependency parsing and graph convolutional network layers to attempt to achieve high accuracy on resolving Winograd problems. Used the model to participate in the Winogrande challenge. Our results as of writing our paper placed us in the top 20 of the leaderboards.

The paper can be found in the repo as paper.pdf.

Data

./data/
├── train_[xs,s,m,l,xl].jsonl          # training set with differnt sizes
├── train_[xs,s,m,l,xl]-labels.lst     # answer labels for training sets
├── dev.jsonl                          # development set
├── dev-labels.lst                     # answer labels for development set
├── test.jsonl                         # test set
├── sample-submissions-labels.lst      # example submission file for leaderboard    
└── eval.py                            # evaluation script

Trained a model which employs a RoBERTa with a dependency parsing and graph convolutional network layer to attempt to achieve high accuracy on resolving Winograd problems. Used the model to participate in the Winogrande challenge. Our results as of writing our paper placed us in the top 20 of the leaderboards.

The paper can be found in the repo as paper.pdf