AistrighNLP - A collection of NLP tools for Irish
AistrighNLP is a collection of tools and models used for Aistrigh, the BT Young Scientist 2021 project. Our aim is to bring Irish into the modern era with NLP tools to give it parity with English. The tools included are based around the work in Neural Models for Predicting Celtic Mutations (Scannell, 2020). Included is all the tools needed to create a demutated Irish corpus, which can be used in all sorts of NLP tasks, and a model to reinsert them. For the Python API docs visit AistrighNLP Python API
Installing the Package
AistrighNLP can be downloaded using
pip install aistrigh-nlp
When lowercasing either Irish or Scots Gaelic for prediciting mutations, you must be aware of special cases outlined in the paper above. Our lowercaser handles that
aistrigh-nlp lowercase -i input.txt -o output.txt
To remove mutations from an entire dataset for use for NLP tasks (like Machine Translation) use
--lang must take ISO 639 language codes like 'ga'.
aistrigh-nlp demutate-corpus -i input.txt -o output.txt -l ga
To remove mutations with a 'window' on either side to train a neural network, use
-w set to your desired window length on each side
aistrigh-nlp demutate-window -i input.txt -o output.csv -l ga -w 16
Predicting the mutations
To predict mutations on each word, use
predict-mutations. As of right now, it's only compatible with PyTorch+Torchtext models but we are working on expanding to TensorFlow and Keras. You'll need your vocab, labels and model checkpoint in the same folder (
--data). We provide default models to be used here.
aistrigh-nlp predict-mutations -i input.txt -o output.txt -w 16 -d nn_100k
Applying the predicted mutations
To apply the mutations predicted by
aistrigh-nlp apply-mutations -i input.txt -o output.txt -l ga
Scoring Celtic NMT models using Standard and Demutated BLEU
To score NMT models using both these metrics run;
aistrigh-nlp bleu -r reference.txt -p predictions.txt -l ga
If you're scoring a demutated NMT model and haven't reapplied mutations, pass your demutated reference (
-d) and predictions, and the original reference (
aistrigh-nlp bleu -d demutated_reference.txt -r reference.txt -p predictions.txt -l ga
AistrighNLP uses PyTorch
Traces to save the full computational graphs as checkpoints. This way, the model architecture need not be declared into hard-coded scripts. See this StackOverflow Thread for instructions to save a traced checkpoint.
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.