I just wanted to share a suggestion and see if the developers think it’s interesting. There is currently a very nice pytorch implementation of tree-structured LSTM networks here https://github.com/dasguptar/treelstm.pytorch
I was wondering if you think it would make for an interesting addition to FastAI. In a way, in my untrained eyes, it does look similar to the language model approach, taking into account syntantic position of words. Now, I would guess that this being a slightly old idea (2015 paper here https://arxiv.org/abs/1503.00075) it isn’t necessarily better than the language model approach in [sentiment] classification but perhaps the ability of assessing relatedness of two input sentences might be interesting for some people.