Implementing "A Deep Reinforcement Model for Abstractive Summarization" with OpenNMT-py

Objectives, challenges and todo

Posted on October 23, 2017

Model overview The paper: https://arxiv.org/pdf/1705.04304.pdf Encoder Shared Embeddings (dim=100), vocab src=150k, tgt=50k Bi-directionnal LSTM (dim=2x200) Decoder LSTM (dim=400) Encoder Temporal Attention (between $ h^d_t \text{ and } h_e^t $), producing attention scores ($a^e_t$) and context vector ($c^e_t$) Decoder Attention (between $h^d_t$ and $h^t_{t-1}$... Read More