Because of their superior ability to preserve sequence information over time,\nLong Short-Term Memory (LSTM) networks, a type of recurrent neural network with\na more complex computational unit, have obtained strong results on a variety of\nsequence modeling tasks. The only underlying LSTM structure that has been\nexplored so far is a linear chain. However, natural language exhibits syntactic\nproperties that would naturally combine words to phrases. We introduce the\nTree-LSTM, a generalization of LSTMs to tree-structured network topologies.\nTree-LSTMs outperform all existing systems and strong LSTM baselines on two\ntasks: predicting the semantic relatedness of two sentences (SemEval 2014, Task\n1) and sentiment classification (Stanford Sentiment Treebank).\n