A C++ implementation of the long short-term memory recursive neural network model (LSTM-RNN) described in

[1] Phong Le and Willem Zuidema (2015). Compositional Distributional Semantics with Long Short Term Memory. In Proceedings of Joint Conference on Lexical and Computational Semantics (*SEM).

Written and maintained by Phong Le (p.le [at] uva.nl)

Note: this is similar to Tree-LSTM, which was proposed at the same time.

This package contains three components

• src - source code files in C++ of the LSTM-RNN,

• data - Stanford Sentiment Treebank and GloVe word embeddings

• Release - for compiling the source code.

• Install OpenBlas at /opt/OpenBLAS.
• Go to Release, execute make. It should work with gcc 4.9 or later. (If OpenBlas is not installed at /opt/OpenBLAS, you need to replace /opt/OpenBLAS by the correct path in makefile and src/subdir.mk.)

The following instruction is for replicating the second experiment reported in [1]. Some small changes are needed for your own cases.

data/trees contains the train/dev/test files of the Stanford Sentiment Treebank (SST).

data/dic/glove-300d-840B contains the 300-D GloVe word embeddings (vectors of words not in the SST are removed).

If you want to use other word-embeddings, you should follow the following format:

• create words.lst containing all frequent words (e.g. words appear at least 2 times), each word per line. The first line is always #UNKNOWN#
• create wembs.txt containing word vectors. The first line is <number-of-vectors> <dimension>. For each following line: <word> <vector>. Vectors of words in words.lst but not in wembs.txt will be randomly initialised.

In Release, open train.config, which stores the default parameter values. It should look like

dim 50

lambda 1e-3
lambdaL 1e-3
lambdaC 1e-3
dropoutRate 0
normGradThresh 1e10

learningRateDecay 0.
paramLearningRate 0.01
wembLearningRate 0.01
classLearningRate 0.01
evalDevStep 1

maxNEpoch 10
batchSize 5
nThreads 1

dataDir ../data/trees/
dicDir ../data/dic/glove-300d-840B/

compositionType LSTM
functionType tanh

You can try the traditional RNN by setting compositionType NORMAL, and try other activation functions by setting functionType <NAME> where <NAME> could be softsign, sigmoid, rlu.

Execute

./lstm-rnn --train train.config model.dat

where the resulting model will be stored in model.dat.

Execute

./lstm-rnn --test <test-file> <model-file>