JS Code Completion as a "N"LP Problem

js-code-completion-nlp

JavaScript Code Completion as a "N"LP Problem
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

http://courses.csail.mit.edu/6.864/fall_2016/past_projects.html

Many techniques in natural language processing rely on statistical 
methods because natural language is fuzzy and imprecise. In contrast, 
computer languages follow precise known rules, so tools for working 
with computer languages are usually strictly rule-based. 

However, even in the realm of computer languages, some analysis tasks 
may benefit from statistical methods by being able to capture 
regularities in how computer code is authored, such as in the usage 
of common APIs. 

In this project I look at the task of code completion for traditional 
JavaScript, a language that presents some challenges for static analysis 
in the traditional sense due to its dynamic nature. 

The task is as follows: given a corpus of existing JavaScript code and 
a test program that includes blanks to be filled in, suggest a list 
of possible completions for each blank. 

I develop several approaches for this task, including a discriminative 
model for predicting attribute and method names using contextual features 
based on the abstract syntax tree of the JavaScript program.