how to implement tree-lstm proposed in keras and how the tree structure is constructed?

I want to implement the dependency tree-lstm proposed in https://arxiv.org/pdf/1503.00075 for sentiment classification in Keras, but I do not understand how the tree structure is constructed. The paper does not mention the construction procedure of the tree. Does the model itself somehow learn to construct the tree for each textual input data? and isn't there any Keras implementation of tree-lstm in Keras?