thesis on using ML for Pharo code completion

I'm starting work on a thesis on using ML for Pharo completion, in particular n-gram language models, and I wrote a little blog post with some details and research questions that I have. Here's a link: https://medium.com/@myroslavarm/using-ml-for-pharo-autocompletion-ideas-80a362258bc3.

Any feedback or ideas on the topic (and research questions) are most welcome.

Best regards,

Myroslava

Hi Myroslava,

I read your post (actually a few of them) and I have some questions:

- I was surprised to read about n-gram model because at least in NLP was outperformed by modern DL approaches. But then I read you considered RNN (and maybe attention model?) so it sounds really interesting.

- Do you plan to benchmark agains TabNine? I know they are a black box, but they seem to support Smalltalk https://github.com/zxqfl/TabNine/blob/master/languages.yml

- ERNIE 2.0 supports continual learning (AFAIK this was only applied to NLP) so develop something which learns incrementally would be an option for your PhD?

One problem, for me, of code prediction is gathering enough context information to make informed predictions (I ignore if there is something like GLUE benchmarks for code). This means to propose contextualized completions based on multiple factors: for example the class where I am positioned in the Browser, the recent used/written methods, etc. Of course the possibilities are infinite! Just thinking out loud.

Cheers,

Hernán