Part-of-Speech Tagging for Icelandic using a Bidirectional Long Short-Term Memory Model with combined Word and Character embeddings
A bidirectional LSTM model can be trained with the script train.py.
Running ./train.py -h gives information on all possible parameters.
The program requires input corpora to be in the same format as the IFD-training/testing sets, available at http://www.malfong.is/index.php?lang=en&pg=ordtidnibok
Results from experiments with different models, trained on the IFD, are here: https://docs.google.com/spreadsheets/d/1YG8xYaW10x4jvsCFyHkQxdYQLIOFa6WoBUzbA4y4Gtg/edit?usp=sharing
During long running training sessions get_results.py was used fetch the latest results and calculate the highest scoring epoch.
If the parameter -p is added when running the script a plot is provided for the accuracy and avg. loss of all epochs finished sofar.
The script interactive.py trains a model and then allow the user to try out the tagger on sentences he enters. It accepts the same parameters as train.py.
