Learning Python Code Suggestion with a Sparse Pointer Network

• DOI: 10.48550/arxiv.1611.08307
• Publication Date: 2016-01-01
• Authors: Avishkar Bhoopchand; Tim Rocktäschel; Earl T. Barr; Sebastian Riedel
• Keywords: Python

To enhance developer productivity, all modern integrated development environments (IDEs) include code suggestion functionality that proposes likely next tokens at the cursor. While current IDEs work well for statically-typed languages, their reliance on type annotations means they do not provide same level of support dynamic programming languages as languages. Moreover, engines in propose expressions or multi-statement idiomatic code. Recent has shown language models can improve systems by learning from software repositories. This paper introduces a neural model with sparse pointer network aimed capturing very long-range dependencies. We release large-scale corpus 41M lines Python crawled GitHub. On this corpus, we found standard to perform suggesting local phenomena, but struggle refer identifiers are introduced many past. By augmenting specialized referring predefined classes identifiers, obtain much lower perplexity and 5 percentage points increase accuracy compared an LSTM baseline. In fact, is due 13 times more accurate prediction identifiers. Furthermore, qualitative analysis shows indeed captures interesting dependencies, like class member defined over 60