FIESTA (Fast IdEntification of State-of-The-Art models using adaptive bandit algorithms)

Quick links:

1. Documentation - You can find the motivation of the project code base there as well.
2. Tutorials

Installing

Requires Python 3.6.1 or greater.

pip install fiesta-nlp

Experiments in the paper

NER experiments

The code used to create the NER results can be founder here with all of the instructions on:

1. How the data was split.
2. How to re-run the models.
3. How the images in the paper were created.
4. Links to all of the original F1 results and data splits.

Target Dependent Sentiment Analysis experiments

The 500 Macro F1 results from the 12 different TDSA models can be found within test_f1.json file. For replication purposes we have created a Google Colab notebook which can be found here that shows how the results from the paper can be replicated. Further more this notebook is a good example of how to use the fiesta library when you already have results and do not need to evaluate any modles.

Citing (This will be updated when the ACL version of the paper is published)

If you use FIESTA in your research, please cite FIESTA: Fast IdEntification of State-of-The-Art models using adaptive bandit algorithms

@article{moss2019fiesta,
  title={FIESTA: Fast IdEntification of State-of-The-Art models using adaptive bandit algorithms},
  author={Moss, Henry B and Moore, Andrew and Leslie, David S and Rayson, Paul},
  journal={arXiv preprint arXiv:1906.12230},
  year={2019}
}

General Acknowledgments

This code base and it's related FIESTA paper could not have been done without:

1. Henry Moss's time funded through EPSRC Doctoral Training Grant and the STOR-i Centre for Doctoral Training.
2. Andrew Moore's time funded through EPSRC Doctoral Training Grant.
3. Paul Rayson's and David Leslie's time.
4. Resources -- The loan of a NVIDIA GP100-equipped workstation from Dr Chris Jewell at the Centre for Health Informatics, Computing, and Statistics, Lancaster University.
5. We lastly thank the comments and advise of the reviewers from ACL 2019 which has greatly improved the paper.

Issue template Acknowledgment

We copied/adapted the issues templates from the allennlp project.