I am doing PhD in AI field mainly in NLP

my supervisor told that I have to contribute something new to the field.

OK, I started doing some trails in improving one case in the field of NLP AI

I found a way that can improve that case in 80% of the times.

I don't want to go into the details of the method that I found, as it is mainly technical.

Anyway, I told my supervisor about what I found and he wasn't impressed.

The things he mentioned that this method is not provable, which means it is proved only by trying the cases without mathimatical prove.

is this enough for PhD or shall I prove it mathematically?

  • 4
    Are you sure he meant a mathematical proof? If so, are you working in a more theory-oriented group? For an NLP "advance," I usually expect to see good results on a benchmark problem.
    – cag51
    May 2, 2021 at 3:13
  • @cag51 what do you mean by theory oriented group?
    – asmgx
    May 2, 2021 at 4:20

2 Answers 2


Let me add to the answer of Bryan Krause.

First, a technique that is an "improvement 80% of the time" might be extremely dangerous to apply in practice. Even proving that the quoted phrase is true may not be enough.

What you would need to do, in my view, is determine the parameters that lead to its assured success vs its failure. That analysis is probably worth more than a proof, per se. Why is it sometimes better and why is it sometimes worse.

And of course I'm assuming that you have a good definition of "improvement".

Note that some AI/NLP practices have led to race-biased outcomes among other difficulties.


Your thesis committee, and your supervisor in particular, are the ones who decide if your research is sufficiently novel.

Make sure they are happy and seek their guidance if they are not. It seems your supervisor isn't happy, so work with them to find what you can add or what other direction you can take. You could provide a StackExchange answer with dozens of upvotes saying your work is a novel contribution and it won't get you closer to graduating.

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .