Electra-AGE_FE

Code and splits for the paper "Combining ELECTRA and Adaptive Graph Encoding for Frame Identification" at LREC2022.

The code is in this repository, while the splits must be downloaded from this link and extracted in the repository root folder.

For reproducing the experiments:

1. Create ELECTRA embeddings for a given split (here FN1.5_08):

    cd scripts
    bash ./CreateVectors.sh 1.5_08

2. Perform the entire set of 10 experiments with different random initialisation:

    bash ./FrameID.sh -m train -x 1.5_08 -v 1.5 -e electra 

3. Look at the final detailed results on the test set:

    python3 ./showResults.py ../experiments/xp_1.5_08/frameid/results/summary

Results obtained on a NVIDIA Titan XP 12GB and:

• CUDA 10.0
• cuDNN 7.4.2
• python 3.6.7
• pytorch 1.4.1
• transformers 3.1.0
• networkx 2.2

Parts of this package are based on pyfn.

If you find this code useful in your research, please cite:

@InProceedings{Tamburini:2022:LREC,
  author    = {Tamburini, Fabio},
  title     = {Combining ELECTRA and Adaptive Graph Encoding for Frame Identification},
  booktitle = {Proceedings of The 13th Language Resources and Evaluation Conference},
  month     = {June},
  year      = {2022},
  address   = {Marseille, France},
  publisher = {European Language Resources Association},
}

In case of problems contact me at fabio.tamburini@unibo.it.