An attempt to implement the Constinuency Tree LSTM in "Improved Semantic Representations From Tree-Structured Long Short-Term Memory Networks" (Tai, Socher, and Manning, 2015) in Theano.

• Python 2.7
• NumPy
• Theano
• gensim

• Stanford Sentiment Treebank in PTB Tree format: download
• pretrained word2vec word vectors trained on Google News, available here

• download the source files or clone the repo
• download the Stanford Sentiment Treebank in PTB format using the above link
• download the pre-trained word2vec vectors using the above link, and unzip the file

To run using default settings:

python constituency_tree_lstm.py /path/to/stanfordSentimentTreebank/trees --word2vec_file=/path/to/GoogleNews-vectors-negative300.bin

To see additional options:

python constituency_tree_lstm.py -h

The original Constituency Tree LSTM was implemented in Torch. In addition, the following are different about this implementation

• no minibatches
• adjusted learning rate and regularization strength to compensate for minibatches of size=1
• using word2vec vectors, instead of Glove
• no updatting of embeddings during training

Performance is not yet equivalent to the results of the original paper, perhaps because of the above differences. The best performance I have been able to obtain thus far is ~0.85 on the binary classification task.