Scheduled Sampling

[anirudh9119]

Results look promising to me, http://arxiv.org/pdf/1506.03099v3.pdf + the future direction as pointed by the authors. Any thoughts/discussions you had regarding this to NMT?

[bartvm]

I personally like the idea because I think it addresses an important problem. That said, some people I talked to about this pointed out that the scores they give in the paper are for image captioning, and aren't that great. For machine translation the language generation is more constrained (there's less freedom with translating than with captioning) so the benefit is probably less strong.

On the other hand, it can't hurt either, so it's just a question as to whether the (minor) score improvements outweigh the implementation time and complexity.