I need to map from a vector space representation onto a tree structure.
A possible solution: given a word vector as input, produce a path in the tree from the root down to the node that most closely matches that vector.
The path would be a variable length string composed from a finite set of symbols. The variable length output leads me towards long short-term memory (LSTM) models. But I've never built a complete LSTM model before.
The understanding gap: my vector inputs are already dense representations. Specifically, I'm working with GloVe right now. But the output path symbols would require a different encoding, correct? How can I structure/train the necessary encoder-decoder pair so that it can handle both word vectors and these path symbols as input while still being able to produce path symbols as output?
EDIT: After more research, I think the correct term would be a "one to many" model using an RNN architecture (of which LSTM is one type). So I wouldn't need an encoder for the one input vector because it's already a dense vector. But I would need to train an autoencoder for decoding the multiple output vectors that describe the tree path.