This paper checks all the boxes for me. The prior approaches are elegantly tied together, the novel contributions are straightforward and intuitive, the results are compelling, and it is extremely clearly written. Kudos to the authors!
It'd be crazy if an event happened, and you could choose how much you wanted to read about it - "I want to read at least 5 paragraphs on this subject." Could help a lot for textbooks if you really understood a subject (or conversely really needed some help on one.)
Really interesting text-to-text generation work. There’s also some work being looking at using Neural Networks for data-to-text generation. Some of which was presented at INLG 2016 last year at Edinburgh. However, it is still early days and currently symbolic approaches still achieve better text quality in closed world applications.
What do you mean by "data-to-text"?
Very cool! Would be awesome to eventually run the summaries the RNN generates in a production-ready capability, and see how it does. Reddit has a few TLDR bots that I think are more extractive and could use an update.
Maybe a similar idea would be running those bots as evaluation tasks. I suppose voting would be a decent proxy objective when you are dealing with very qualitative results and don't want to spend a ton of time evaluating them yourself. Spin up a bunch of bots with different ideas and see which ones score the best.
Really promising and interesting research, thanks!