nala-cub / g2p-pgt

Code implementation for the Paper "Making a Point: Pointer-Generator Transformers for Disjoint Vocabularies" (https://www.aclweb.org/anthology/2020.aacl-srw.13.pdf).

SIGMORPHON 2020 Shared Tasks

Task 1

This repository is a fork of the vanilla transformer implementation of shijie-wu, providing the extra features of Multi-tasking for G2P (Grapheme-to-Phoneme Conversion) and P2G (Phoneme-to-Grapheme Conversion), and a pointer-generator copy-mechanism.

Dependencies

• python 3
• pytorch==1.4
• numpy
• tqdm
• fire

Install

make

Train Example

python src/train.py --dataset g2p --train data/kor_train.tsv --test data/kor_test.tsv --dev data/kor_dev.tsv --arch transformer --model model/kor --opt_pickles model/kor-opt-pickle --align data/kor_alignment.pkl --use_copy True --seed 12

Flags

--use_copy [True|False]   use pointer-generator mechanism
--opt_pickles [path]    path to pickle file storing performance results including WER and PER
--align [path]    path to alignment file (which we generated using GIZA++)
--freq_eval_after [INT]    Epoch after which evaluation starts
--freq_eval_every [INT]    Frequency of evaluation after freq_eval_after

License

MIT