g2p2022

This repo has the relevant files to recreate our entry for the SIGMORPHON 2022 G2P shared task.

Three systems

We report three different systems in the paper.

1. "vanilla" transformer run on CMU-dict English data. This is the file vanilla.py.
2. OpenNMT transformer model run on the shared task data. The configuration files are located in the nmt directory. This model was run in a docker container and instructions for creating the image are below.
3. Phonetisaurus model run on the shared task data. Configuration files are in the phone directory. This model was also run in a docker container and instructions for creating the container are below.

OpenNMT container

First, create a docker container:

docker run -it \
  --gpus all \
  --name nmt \
  -v /data/:/mhdata \
	-v /home/mhammond/sigmorphon2022:/mhsig \
	pytorch/pytorch:1.7.0-cuda11.0-cudnn8-runtime

Run these steps:

apt update
apt upgrade
apt install vim
apt install wget

Then install:

pip install OpenNMT-py

Phonetisaurus container

First, create a docker container:

docker run -it \
  --gpus all \
  -p 8888:8888 \
  --name phone \
  -v /data/:/mhdata \
  -v /home/mhammond/sigmorphon2022/:/mhsig \
  ubuntu:20.04

Next clone the Phonetisaurus repo:

https://github.com/AdolfVonKleist/Phonetisaurus

Do the steps from the readme.md file there adding these bits:

apt install wget
apt install python3-pip
apt install vim
...
vim test.wlist