Neural Code Synthesis and Completion

In my MS thesis, I presented my work on Neural Code generation from natural language and Neural code completion.

Goal

• Natural language to code translation
  • Aims to develop a versatile Seq2Seq architecture for both objectives of translating text to code (NL2Code) and of generating comments, docstring,
    method documentation from source code input (Code2NL).
  • Aims to use various neural-based subword tokenizers to incorporate the contextual embeddings of the input.
  • Aims at performing an ablation study to gauge the importance of the crucial components of the developed AI system.
  • Aims to develop transfer learning and data augmentation techniques to generate more diverse and accurate source code translations.
• Code completion
  • Aims to develop a novel RoBERTa based neural language model for source code.
  • Aims at investing the performance of the model for the fill-in-mask task and compare the predicted masked tokens with the ground truth.