conv-attention-network-source-code-summarization.md

A Convolutional Attention Network for Extreme Summarization of Source Code

• Arxiv link
• Source code
• Data set

Application

• This conv + attention network learns to summarise source code functions
• Given a sequence of words from the source code, outputs a sequence for the suggested method name

Summary

• 1D convolution of source code tokens feeds into 2 layers of stacked attention network
• Also incorporates a copy framework to suggest out of vocabulary tokens
• Weighed sum of the attention tokens is decoded via GRUs and hybrid beam and BFS
• Data is collected from 10MB of Java repos on Github

(Attention visualization for getNode())