CodeFill: Multi-token Code Completion by Jointly Learning from Structure and Naming Sequences

TL;DR

CodeFill is a novel code autocompletion model that jointly learns from code structure and naming sequences, enabling more accurate multi-token and single-token predictions by capturing long-range dependencies.

Contribution

The paper introduces CodeFill, a multi-task Transformer model that combines structure and naming information for improved code autocompletion, trained on large datasets and evaluated with realistic benchmarks.

Findings

Outperforms baselines in single-token prediction (MRR: 70.9%)

Achieves state-of-the-art multi-token prediction (ROUGE-L: 63.7%)

Effectively captures long-range dependencies in code

Abstract

Code completion is an essential feature of IDEs, yet current autocompleters are restricted to either grammar-based or NLP-based single token completions. Both approaches have significant drawbacks: grammar-based autocompletion is restricted in dynamically-typed language environments, whereas NLP-based autocompleters struggle to understand the semantics of the programming language and the developer's code context. In this work, we present CodeFill, a language model for autocompletion that combines learned structure and naming information. Using a parallel Transformer architecture and multi-task learning, CodeFill consumes sequences of source code token names and their equivalent AST token types. Uniquely, CodeFill is trained both for single-token and multi-token (statement) prediction, which enables it to learn long-range dependencies among grammatical and naming elements. We train CodeFill on two datasets, consisting of 29M and 425M lines of code, respectively. To make the evaluation more realistic, we develop a method to automatically infer points in the source code at which completion matters. We compare CodeFill against four baselines and two state-of-the-art models, GPT-C and TravTrans+.CodeFill surpasses all baselines in single token prediction (MRR: 70.9% vs. 66.2% and 67.8%) and outperforms the state of the art for multi-token prediction (ROUGE-L: 63.7% vs. 52.4% and 59.2%, for n=4 tokens). We publicly release our source code and datasets.