NLPLingo

BBN's Deep Learning toolkit for event extraction and causal relation extraction.

Coding conventions

These are minimal, as follows:

• We use PyCharm as our UI for coding. Set your tab stops to 4 spaces.
• Class names should be mixed case, e.g. EventExtractionModel
• Variable names, function names, should be lower case, e.g. calculate_accuracy
• Although Python is free type, but adding type hints in your code allows code completion, and is useful to whoever is reading your code, e.g.

def add_entity_mention(self, entity_mention):
    """:type entity_mention: nlplingo.text.text_span.EntityMention"""

You can also add type hints after variable declaration, e.g.:

self.entity_mentions = []
""":type: list[nlplingo.text.text_span.EntityMention]"""

Setting up nlplingo

nlpLingo now uses Tensorflow as backend.

Setting up tensorflow GPU anaconda:

conda create --name p3-gpu
source activate p3-gpu
conda install -c anaconda python=3.6.8
conda install -c anaconda tensorflow-gpu=1.12 (use 'tensorflow=1.12' for CPU)
conda install keras=2.2.4
conda install -c anaconda future
conda install spacy
conda install -c anaconda simplejson (else you will get an error "simplejson has no attribute loads" when you try to do the 'spacy download')
python -m spacy download en
conda install -c anaconda pytorch (use '-c anaconda pytorch-cpu' for CPU)
conda install -c conda-forge python-crfsuite (optional, just for NER)

Note: tensorflow seems to need GLIBC 2.12, while torch needs GLIBC 2.14. To add GLIBC 2.14: export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/glibc-2.14/lib But if you do that, then tensorflow will not work, while pytorch will work

Testing that your GPU install is working:

import tensorflow as tf
tf.test.is_gpu_available()

import torch
torch.cuda.is_available()

Do export KERAS_BACKEND=tensorflow

Or set here permanently /nfs/raid84/u12/ychan/anaconda2/envs/p3-gpu/etc/conda/activate.d/keras_activate.sh

If you want to add the python env to your jupyter kernel:

conda install -c anaconda ipykernel
conda install -c anaconda jupyter
ipython kernel install --name p3-gpu --user
jupyter kernelspec list (shows the python kernels that you can use for jupyter)

The following that uses Theano as the backend is now deprecated.

conda create --name nlplingo python=2.7.12 numpy=1.12.1 scipy=0.19.0 openblas=0.2.19 theano=0.9.0 keras=2.0.2 spacy=1.8.2
source activate nlplingo
python -m spacy download en
conda install -c anaconda future
conda install -c anaconda scikit-learn
conda install -c conda-forge jsonpickle
conda install -c conda-forge python-crfsuite
conda install -c anaconda mkl-service

export env KERAS_BACKEND=theano
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/cuda-8.0/lib64
export CUDA_ROOT=/opt/cuda-8.0

In ~/.keras/keras.json , set "backend" : "theano"

Running nlplingo

You need to point PYTHONPATH to 2 locations:

• dir that contains serifxml.py
• dir where nlplingo resides

This is an example command: PYTHONPATH=/home/ychan/anaconda2/envs/bbn_env/lib/python2.7/site-packages/serif:/home/ychan/repos/nlplingo python /home/ychan/repos/nlplingo/nlplingo/event/train_test.py --params x.params --mode test_argument

Here are the available modes for train_test.py:

• train_trigger
• test_trigger
• train_argument
• test_argument
• decode_trigger_argument

Comments:

• The scores are written to *.score files
• The dataset is divided into train, dev, test. train_.score files report evaluation on dev set. test_.score files report evaluation on test set.
• We'll normally do hyper-parameters search over (cnn.neighbor_dist, *.positive_weight, *.batch_size, *.num_feature_maps, *.epoch)
• Note that each run will produce different scores/results (e.g. due to drop-out). So we also normally do a few runs per hyper-params setting, and take the one which performs best on dev.
• The hdf and pickle are model files.

Acknowledgments

This work was supported by DARPA/I2O and U.S. Air Force Research Laboratory Contract No. FA8650-17-C-7716 under the Causal Exploration program, and DARPA/I2O and U.S. Army Research Office Contract No. W911NF-18-C-0003 under the World Modelers program. The views, opinions, and/or findings expressed are those of the author(s) and should not be interpreted as representing the official views or policies of the Department of Defense or the U.S. Government. This document does not contain technology or technical data controlled under either the U.S. International Traffic in Arms Regulations or the U.S. Export Administration Regulations.