A Deep Learning Predictive Model for Source Code Analysis of a Dynamically Typed Language

The purpose of this study is to use the concepts learned in NLP (Natural Language Processing), combined with machine learning techniques to analyze a computer programming language. Given the many applications and usefulness of NLP methods for plain text, the programming community is very much interested in how similar studies or investigations can be carried on computer languages and eventually simplify source code writing or improve their readability. Our study aims to generate text code suggestion based on cursor input. Code completion is a context-aware process that could speed up the task of coding applications. We first consider this study for dynamically typed programming language mainly because most studies in this field are carried out on statically typed languages like C++ and java. This study could have significant application in code generation, source completion and source code maintenance tasks. We build a deep learning model based on LSTM network using keras (a neural network API) with a Tensorflow backend. Tensorflow is an open source library for numerical computation developed by researchers of the Google Brain Team. We evaluate our model on a large corpus of source code harvested from an open source projects repository. The Concrete Syntax Tree (CST) of the source code is used to capture the semantics and structure of the source code and to increase the repetitiveness of the tokens in the dataset corpus. The performance of our models will be compared to similar studies conducted in the field