FIDO

This is the code for our paper Exploiting Definitions for Frame Identification (EACL 2021).

Testing Environment

This project is built on python==3.7.6, torch=1.4.0, transformers==2.9.0.

Use Our Pre-trained Models - FIDO

Download the pre-trained models at: trained on FrameNet 1.5, trained on FrameNet 1.7.

The accuracy of the pre-trained models are:

	dev	test
FN 1.5	92.4	91.5
FN 1.7	92.4	92.3

The file you would like to predict should be "data/fn1.5/test.csv" or "data/fn1.7/test.csv".

Put the extracted "model_fn1.5" or "model_fn1.7" folder under the "pretrained_models/" directory, and run predict.sh.

It will generate two files: "test_prediction_labels.txt" and "test_prediction_probs.txt" under the model directory.

Train from Scratch

For FrameNet 1.5, data files (train.csv, dev.csv and test.csv) should be put under "data/fn1.5/".

For FrameNet 1.7, similarly, data files should be put under "data/fn1.7/".

Run train.sh. You should get similar results compared to the table above.

Data Format

id, sentence, lu_name, lu_head_position, lu_defs, frame_names, frame_defs, label

lu_name: the target word or phrase
lu_head_position: the position index of the target in the sentence
lu_defs: all the target word definitions associated with the candidate frames (each LU will have different definitions for different associated frames), separated by "~$~"
frame_names: candidate frames, separated by "~$~"
frame_defs: candidate frame definitions, separated by "~$~"
label: an integer indicating the correct frame from the frame_names (lu_defs, frame_names and frame_defs should have the same corresponding order)

An example of the data format can be found under "data/fn1.5/".

The "data/fn1.5/" folder *only contains a small sample* of the data. To replicate the results in the paper, you will need full text of the FrameNet data, as well as the same data split.

Get full access to the FrameNet data at here.

We followed the same train/dev/test split as in Das et al. (2014) and Swayamdipta et al. (2017). Details of the data processing can be found at Open-SESAME.

Contact and Reference

For questions and issues, please contact tianyu@cs.utah.edu. Our paper can be cited as:

@inproceedings{jiang-riloff-2021-exploiting,
title="{Exploiting Definitions for Frame Identification}",
author={Jiang, Tianyu and Riloff, Ellen},
booktitle={Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2021)},
year={2021}
}