N-grams is the first thing you try when you want to statistically model code. We tried it in 2014 and the results are disappointing, if you want to do anything beyond identifying low level syntactic patterns.
The intuition here is that in natural language, context is defined locally. If you want to know whether a 'they' refers to a male or female, for example, you look at the nearby text. In contrast, the flow of data and control through code is highly non-local, which is a lot of why techniques like N-grams don't work.
This has been a very active area of research in academia since we started Kite in 2014. (Suggested google searches: [big code], [ml on code], etc.)
All that said I don't have data on how our approach would compare to N-grams, but I'm guessing if you look at some of the academic research, you'll find that NLP techniques were abandoned, despite the early papers' focus on them.
The intuition here is that in natural language, context is defined locally. If you want to know whether a 'they' refers to a male or female, for example, you look at the nearby text. In contrast, the flow of data and control through code is highly non-local, which is a lot of why techniques like N-grams don't work.
This has been a very active area of research in academia since we started Kite in 2014. (Suggested google searches: [big code], [ml on code], etc.)
All that said I don't have data on how our approach would compare to N-grams, but I'm guessing if you look at some of the academic research, you'll find that NLP techniques were abandoned, despite the early papers' focus on them.