• See the .ipynb files (viewable on Github) for instructions and demo using the script

• Intent classification: intents_train.ipynb intents_predict.ipynb

• Entities recognition: entities_train.ipynb entities_predict.ipynb

• Complete the environmental setup to run the .ipynb files

1. Run data/get_data.bash to download data files Notes:

• Modify this file to selectively download only required files (usually either GLoVE/fasttext vectors)
• You might need to run chmod u+x data/get_data.bash as well to grant execute permissions
• Technically, the models can be generalized to use vectors from https://github.com/facebookresearch/fastText/blob/master/docs/crawl-vectors.md for all other languages

Data path configurations are stored in config.py

• The project (might) also uses 20.000 most common words list from https://github.com/first20hours/google-10000-english/blob/master/20k.txt

2. Setup Python environment:

• Download Anaconda/Miniconda for Python 3.6
• Create a new environment for the project by running conda create --name botbot-nlp python=3.6
• Switch to the newly created environment by running source activate botbot-nlp (activate botbot-nlp inside Anaconda Prompt on Windows)
• Install dependencies by running pip install -r requirements.txt or pip install -r requirements_win.txt for Windows
• Install any missing dependencies (because of platform differences)

3. Build Cython modules

• Run python setup.py build_ext --inplace (requires gcc/clang - sudo apt-get install build-essential on Deb/Ubuntu or Xcode+CLI tools on OSX)

4. Using Jupyter notebook for evaluation

• Activate the environment by source activate botbot-nlp
• Navigate to the root directory by cd
• Run jupyter notebook. A browser tab should open and navigate to jupyter notebook at localhost:8888 by default
• Open entities_train.ipynb inside the notebook
• Click Kernel > Run All to start training

Note: if the progress bars doesn't show up properly during training, run conda install -c conda-forge ipywidgets

train_quora.py trains the paraphrasing model on the Quora duplicate questions dataset

train_sent_to_vec.py trains the InferSent model on NLI + SNLI

train_amazon_sentiment.py trains the classification model on the amazon sentiment dataset

train_conll_eval trains the entity recognition model on the CoNLL2003 dataset

start_flask.py (to be used with -debug True) is for debugging the NLU flask server

The model should be run on Gunicorn by

gunicorn -w 1 -t 0 -b 127.0.0.1:5000 flask_app.entrypoint:app

Arguments:

• -w number of workers
• -t timeout (set to a high number because loading word vectors takes a while)
• -b optionally binds to a different address

After running the server

1. /upload (POST) is used to upload a data file & train the NLU on the data file with the file argument - e.g: curl -X POST -F 'file=@./data/botbot/francis.json' 127.0.0.1:5000/upload

curl -X POST -F "file=@./data/botbot/francis.json" 127.0.0.1:5000/upload

Returns: model_id

2. /predict (POST) is used to send a query for prediction

e.g

curl -X POST -H "Content-Type: application/json"  -d '{"query":"Hello world!", "model_id": "guid_in_previous_step"}' 127.0.0.1:5000/predict | json_pp

1. Using GPU

Using GPU will massively speed up training and inference time (brings training from hours of CPU time to about an hour or a few minutes depending on GPU spec)

e.g:

• Training on CPU: Intel Core i7-4710HQ @2.7Ghz: ~45m/epoch - 12 epochs to reach 87% accuracy

• Training on GPU: same machine, NVIDIA GeForce GTX 850M: ~4m/epoch

• Follow Tensorflow-GPU setup instructions at https://www.tensorflow.org/install/install_linux (Including installing the exact CUDA & CuDNN versions - e.g if using Tensorflow 1.5.0 then CUDA 9.0 and CuDNN v7.0 is required even if newer versions exist)

• Run source activate botbot-nlp

• Use the commands for specific platforms on http://pytorch.org/ to install PyTorch

• (Run pip uninstall tensorflow then pip install --upgrade tensorflow-gpu)

2. Using Tensorboard

• Run tensorboard --logdir=bilstm/logs/
• Navigate to localhost:6006 to see training graphs

(These steps are completely optional but are used for deprecated code paths / might help with experimenting)

3. Install HDF5 for Keras model saving/loading

• Install HDF5 from https://support.hdfgroup.org/HDF5/ or by using homebrew:
• Instructions using homebrew (on UNIX):

brew tap homebrew/science
brew install hdf5

• Activate the environment by source activate botbot-nlp
• Install h5py by running pip install h5py

4. Download NLTK data:

• Activate the environment by source activate botbot-nlp
• Run python -m nltk.downloader all

Although almost as effective as LSTMs - Salesforce's QRNN is hardly maintained. I needed to create a fork for the following issues:

• cupy & pynvrtc should be optional imports on a machine without NVIDIA GPU (it's supposed to support CPU as well)
• The CUDA code is encoded twice, once by QRNN library code and once more by pynvrtc. I needed to remove the one in pynvrtc.

To install, run pip install cupy pynvrtc git+https://github.com/luungoc2005/pytorch-qrnn

This demonstrates the use of Facebook fastText for intent classification Entities recognition is a Bi-directional LSTM + CRF implemented in PyTorch This also uses fastText for word embeddings to alleviate the use of Char-CNN + GloVE embeddings for performance (should re-explore the other option some time in the future)

Models:

• Try using spaces as separate tokens (modify bilstm/utils - wordpunct_tokenize function)

Other:

• Setup a docker image for ease getting started
• Setup a small Flask server and CLI for ease in using the project