Character Eyes

Code for our project analyzing character level taggers, accepted at BlackboxNLP 2019.

Contents

• model.py - A fully character level tagger model, implemented in DyNet. It has support for asymmetric bi-directional RNNs, which we found had performance effects depending on linguistic properties of the language.
• Pretrained models for 6 of our 24 languages
• Ready-to-train datasets (from Univseral Dependencies 2.3) for all 24 languages
• This notebook, which produces some of the figures in our paper and gives an overview of our method and metrics. This is the best place to start if you want to use our project.
• An interactive notebook allowing you to visualize the character level activations for our models

Much of the code is modified from Mimick, a character level system that can replace OOVs or UNKs with learned representations approximating a closed vocabulary set of word embeddings.

Citation format

When using our work, please use the following .bib entry:

@article{charactereyes,
  title={Character Eyes: Seeing Language through Character-Level Taggers},
  author={Pinter, Yuval and Marone, Marc and Eisenstein, Jacob},
  journal={arXiv preprint arXiv:1903.05041},
  year={2019}
}