This is an Theano impelmentation of Tree-LSTM for Semantic Textual Similarity(STS) task.
Tree-LSTM is implemented by Torch. More details can be found at https://github.com/stanfordnlp/treelstm .
I would like to compare Torch and Theano for specific task. Therefore, I made this translation. I try three solutions listed below,
-
main_lasagne.py is using Lasagne https://github.com/Lasagne/Lasagne.git
-
main_keras.py is using keras https://github.com/fchollet/keras.git
-
main_theano.py is pure theano implementation with dependency tree structure LSTM
Known issue: both three solutions just get 70% pearson correlation which is so far away from 84% reported in their paper. If you are interested in fix this issue with me, please contact me: jerryli1981@gmail.com
Note that, it is hard to implement dynamic tree stucture LSTM via theano. So I leverage sequence LSTM that already in the Lasagne and keras platforms to predict the relateness score.