Adaptive Code Completion with Meta-learning

Abstract

Since human-written programs have useful local regularities, the ability to adapt to unseen, local context is an important challenge that successful models of source code must overcome. However, the current source code models mostly learn a common code pattern from large scale open-source codebases, which cannot make use of the localness nor satisfy developers’ personal preferences. Consequently, fast learning and adapting to unseen code patterns from limited developers’ code can provide new insights into source code completion. In this work, we train a base code model that is best able to learn semantic and structural information from context to improve predictions of unseen local tokens and propose an adaptive code model leveraging meta-learning techniques. We demonstrate highly improved performance in experiments on a large scale Java GitHub corpus compared with baselines.