A Dead Simple Neural Lemmatizer using AllenNLP

Getting Started

Install AllenNLP with the command

pip install allennlp

To train, run bash.sh, or

allennlp train config/small.json \
    --serialization-dir logs/main \
    --include-package library

All logs will be put in the logs/main directory. If you would like to train again, delete logs/main, otherwise there will be an error.

To predict a trained model, run predict.sh or

allennlp predict logs/main/model.tar.gz data/Bengali_Dataset.txt \
    --output-file data/predict.txt \
    --predictor simple \
    --include-package library \
    --use-dataset-reader

To evaluate a trained model, run evaluate.sh or

allennlp evaluate logs/main/model.tar.gz data/Bengali_Dataset.txt \
    --include-package library

This will output predictions to data/predict.txt.

Visualize Model Performance in TensorBoard

To view tensorboard logs, just run

tensorboard --logdir logs

Make sure tensorflow is installed to be able to run the tensorboard command.

Providing Your Own Data or Model

Prepare your data in the format

word <tab> lemma

with each sentence separated by an empty line. Then split your data into train, validation, and test sets.

Copy a config file in the config directory and change train_data_path, validation_data_path, and test_data_path, to be the paths of your train, validation, and test files respectively.

Then run training with your new config file

allennlp train path_to_config.json \
    --serialization-dir logs/main \
    --include-package library

Finally, output predictions with your new model using

allennlp predict logs/main/model.tar.gz path_to_input_file.txt \
    --output-file path_to_output_file.txt \
    --predictor simple \
    --include-package library \
    --use-dataset-reader