An interesting addition to standard seq2seq + attention models for summarization.
Main additions are a probability of copying a word from the original (allows out of vocabulary words) and a coverage penalty (minimize overlap of current attention vector with sum of previous attention vectors) to avoid repetition.
Looks very promising and has a nice write up (http://www.abigailsee.com/2017/04/16/taming-rnns-for-better-summarization.html) in addition to the paper:
