MLRep4Code Research Project

Source-code-representations for machine-learning -based identification of defective code fragments

...
"The goal of our project is to investigate whether we can combine different approaches to represent source code to find a universal code representation. This representation could be later used by computers and machine-learning algorithms to support code reviewers in finding defective fragments of source code."

Our project in a nutshell

Source-code features under study

  • Metric-based features
  • NLP-based features
  • AST-based features

Tools and models

We use several tools to extract features from code, including our in-home developed tools like CCFlex and ACoRA (BERT4Code).

Programming languages

We study feature extraction from projects written in Java and C/C++.

Code fragments and analysis tasks

We study feature extraction from code at different levels (line, method, file) in the context of different analysis tasks (e.g., coding conventions, security vulnerabilities, code reviews).