Seq2Seq Beam Search Decoding for Pointer Generator

This code is improved from jojonki's repository

The idea of his approach is nice which uses Node and Linked List structure.

But the implementation of Beam Search seems not correct:

• About definition of a step
• And finding end nodes is not correct

I fixed a little bit in his code. And the performance of my Pointer Generator is increased from 3 to 5 points bleu score.

My code can be modified to use in any RNN-based Seq2Seq architecture

Reference:

• https://github.com/jojonki/BeamSearch/tree/master
• http://nlp.seas.harvard.edu/seq2seq-talk/slidescmu.pdf