There is something badly wrong with this paper. Their example shows them restoring μηδέν ἄγαν from μηδέν ἄ??ν. The diacritics are human provided, and very much limit the search space. The inscription to start with would actually be μηδενα??ν (that is, if the number of characters between α and ν is correct, which is also a human provided guess.

All of their other examples seem to start with diacritics as well. What a mistake!

Reminds me of the algorithm from a Canticle for Leibowitz one of the monks was working on. I do have to wonder if things like this do more harm than good. A guess done by deep learning is still just a guess, the difference being the additional danger of it looking more authoritive and certain than it is.

It seems like this particular alogrithm has two main components that are highly important, and human-oriented:

a) The search space is limited with human-supplied options.

b) The output isn't deterministic, is offers a series of best guesses.

I would say this is closer to a NN-powered spellchecker than an authoritative source on the restoration of the work. It still requires a decent amount of interaction from an actual expert for the system to work.

I'm not a DL or ancient greek expert but it seems their NN outputs top20 predictions, so it's more an assistant suggesting characters/words than an authoritative thing. If it helps academicians restoring ancient texts in a way that is faster and with less errors, where is the harm?