A module for E-mail Summarization which uses clustering of skip-thought sentence embeddings.
This code in this repository compliments this Medium article.
- The code is written in Python 2.
- The module uses code of the Skip-Thoughts paper which can be found here. Do:
git clone https://github.com/ryankiros/skip-thoughts
- The code for the skip-thoughts paper uses Theano. Make sure you have Theano installed and GPU acceleration is functional for faster execution.
- Clone this repository and copy the file
email_summarization.py
to the root of the cloned skip-thoughts repository. Do:git clone https://github.com/jatana-research/email-summarization cp email-summarization/email_summarization.py skip-thoughts/
- Install dependencies. Do:
pip install -r email-summarization/requirements.txt python -c 'import nltk; nltk.download("punkt")'
- Download the pre-trained models. The total download size will be of around 5 GB. Do:
mkdir skip-thoughts/models wget -P ./skip-thoughts/models http://www.cs.toronto.edu/~rkiros/models/dictionary.txt wget -P ./skip-thoughts/models http://www.cs.toronto.edu/~rkiros/models/utable.npy wget -P ./skip-thoughts/models http://www.cs.toronto.edu/~rkiros/models/btable.npy wget -P ./skip-thoughts/models http://www.cs.toronto.edu/~rkiros/models/uni_skip.npz wget -P ./skip-thoughts/models http://www.cs.toronto.edu/~rkiros/models/uni_skip.npz.pkl wget -P ./skip-thoughts/models http://www.cs.toronto.edu/~rkiros/models/bi_skip.npz wget -P ./skip-thoughts/models http://www.cs.toronto.edu/~rkiros/models/bi_skip.npz.pkl
- Verify the MD5 hashes of the downloaded files to ensure that the files haven't been corrupted during the download. Do:
The output should be:
md5sum skip-thoughts/models/*
9a15429d694a0e035f9ee1efcb1406f3 bi_skip.npz c9b86840e1dedb05837735d8bf94cee2 bi_skip.npz.pkl 022b5b15f53a84c785e3153a2c383df6 btable.npy 26d8a3e6458500013723b380a4b4b55e dictionary.txt 8eb7c6948001740c3111d71a2fa446c1 uni_skip.npz e1a0ead377877ff3ea5388bb11cfe8d7 uni_skip.npz.pkl 5871cc62fc01b79788c79c219b175617 utable.npy
- Change
Lines:23-24
in the fileskip-thoughts/skipthoughts.py
to provide the correct paths to the downloaded models.path_to_models = 'models/' path_to_tables = 'models/'
- Find any English emails dataset online or create a small one on your own.
- The module expects a list of emails as input and returns a list of summaries.
- Open the Python interpreter in the
skip-thoughts/
folder and do:>>> from email_summarization import summarize >>> summaries = summarize(emails) # emails is a Python list containing English emails.