Reviews: Tree-to-tree Neural Networks for Program Translation

Summary This paper proposes a neural network architecture for tree to tree transduction, with application to program translation. Binarized trees are encoded with a TreeLSTM, and a tree-structured decoder generates binary trees breadth-first, using an attention mechanism over input tree nodes, and attention feeding within the tree decoder. On synthetic and real-world program translation tasks, and across multiple datasets, the proposed model obtains substantially stronger performance than baseline sequential neural models and previous work on automatic program translation. Quality The paper proposes a general framework for neural tree to tree transduction, and the model obtains high performance on program translation tasks. Clarity The model and experimental setup are define clearly.