Building Program Vector Representations for Deep Learning

The main problem when automatically analysing source code is finding a meaningful representation that captures the specific properties of programming languages. While there are widely adopted techniques for building “word embeddings” for natural language (like Word2vec), these are not applicable to source code, whose syntax and vocabulary are simpler .

This paper presents a model for learning “code embeddings”, relying on the syntactic structure of the code, expressed as Abstract Syntax Trees, and how to use those embeddings for code classification tasks.

Paper is available here: http://arxiv.org/abs/1409.3358

Talk will be in english