luofuli / DualRL

A Dual Reinforcement Learning Framework for Unsupervised Text Style Transfer (IJCAI 2019)

https://export.arxiv.org/pdf/1905.10060

Why did you not use beam search for decoding?

jind11 opened this issue 5 years ago · comments

Di Jin commented 5 years ago

I noticed that in your code, you used random and greedy mode for decoding, however, why did you not use beam search? It should be better than greedy decoding at least for evaluation, right? Thanks!

Fuli Luo commented 5 years ago

Greedy decoding is good enough and can get the SOTA results. You can try beam search by just set the decode_type to constants.BEAM at

DualRL/dual_training.py

Line 109 in 2fae5bb

nmt_args.decode_type = constants.GREEDY