Tutorial of AllenNLP at Kording Lab (November, 2018). Here, we implement deep learning model for text classification using AllenNLP. We will go through basic concept on how to implement classifier and how to use the AllenNLP library. All the materials are listed below.

• allennlp_tutorial.pdf is the slide for the lab teaching
• allennlp_tutorial.ipynb is the Jupyter notebook for the AllenNLP example (predict which venue should we submit the paper to)
• venue folder is the folder that we make AllenNLP as a library. To train the model, you can run the following command

allennlp train example_training.json -s output --include-package venue

The file example_training.json contains specification for the training e.g. "cuda_device": -1 means using CPU, "num_epochs": 40 means train for 40 epochs.

You can also see the experiment output via TensorBoard by running

tensorboard --logdir output

More information can be found at allenai/allennlp and allenai/allennlp-as-a-library-example

We do a fun experiment where we train the same model to classify publications from sample 110 journals from MEDLINE. To train the model, you can run the script below. We got an accuracy of 64.9 percent on validation dataset (not too bad!).

allennlp train example_training_110venue.json -s output_110venue --include-package venue

You can see how to use the model in the last section of allennlp_tutorial.ipynb. Alternatively, you can also run the demo locally by downloading the trained model and run allennlp.service.server_simple as follows

wget https://s3-us-west-2.amazonaws.com/allennlp-tutorial/model.tar.gz
python -m allennlp.service.server_simple \
    --archive-path model.tar.gz \
    --predictor venue_predictor \
    --include-package venue \
    --field-name title \
    --field-name paperAbstract \
    --static-dir static_html

This will produce the bar chart output of venues' probability.

Here is a list of dependencies for the demo:

pytorch >= 0.4
spacy
allennlp >= 0.7

You can install these packages using conda and pip commands as follows:

conda install pytorch torchvision -c pytorch
conda install -c conda-forge spacy
pip install "allennlp>=0.7"