Note: This is a modified version of the original code for use in reproducing the original experiments and related research.

This is the cleaned-up code for the paper “Jointly Learning Sentence Embeddings and Syntax with Unsupervised Tree-LSTMs”, currently under review and available at https://arxiv.org/abs/1705.09189. Please refer to http://www.maillard.it/ for the latest information about the paper.

The code for preprocessing the datasets is under data/. The file networks.py contains implementations of the various network architectures. You must select which architecture to use when running the *_train.py/*_eval.py files:

• bow: Bag of Words
• lstm: LSTM
• tree-lstm: Supervised Tree-LSTM
• ltr-cyk: Left-branching Tree-LSTM
• rtl-cyk: Right-branching Tree-LSTM
• rand-cyk: Random-branching non-projective Tree-LSTM
• cyk: Unsupervised Tree-LSTM

Example:

python data/glove2numpy.py \
    ~/data/glove/glove.6B.100d.txt \
    ~/data/glove/glove.vocab.6B.100d.txt \
    ~/data/glove/glove.embeddings.6B.100d.npy