This repo only covers the graph predictor and feature predictor, which are the main parts of 2018 NIPS paper:
Yang, Zhilin, et al. "GLoMo: Unsupervisedly Learned Relational Graphs as Transferable Representations." arXiv preprint arXiv:1806.05662 (2018).
Since the author has not yet release the official code, this repo is only based on personal understanding of the paper, which could be mis-understood or incomplete, we do not recommand a direct usage of this repo at this moment.
This repo will be continually updated, and after the official code released, we will re-implement GLoMo according to its official version ASAP!
Requirements:
python: 3.6
tensorflow: >=1.4.0
TODO:
0.Modification on the implementation of graph predictor and feature predictor
1.Bi-attention mechanism in RNN
2.Downstream task models
3.Training on various tasks