Neural Source Code Descriptor

Code summary, documentation, and retrieval are few of the important uses for creating natural language sequences from source code fragments. Sequence-to-sequence (seq2seq) models, initially developed for neural machine translation (NMT), have achieved amazing success in these tasks by treating source code as a series of tokens. This work introduces code2seq, a novel technique that uses the syntactic structure of programming languages to improve source code encoding. The suggested approach shows a code snippet by breaking it down into a series of compositional paths within its Abstract Syntax Tree (AST) and using attention techniques to pick paths during decoding. We illustrate the efficacy of our approach in the context of code captioning. However, it is worth noting that our model's development is currently at a stage where detailed comparisons with prior research outcomes are limited.