Authors: Hanxiang (Henry) Pan, Xinyu Ma
We approached this problem by using a baseline model that's composed of a gated recurrent autoencoder with attention mechanism.
We analyzed it's performance and pitfalls and implemented the transformer architecture with multi-headed (self) attention mechanism.
We achieved a ROUGE-1 score of 0.35 on the validation set and 0.37 on the training set.
The baseline model can be found in Project_Baseline.ipynb
.
The improved model can be found in Project_Improved.ipynb
.
All the analyses and discussions can be found in Final_Project.pdf
.
MIT license.