Code Editing from Few Exemplars by Adaptive Multi-Extent Composition
This paper considers the computer source code editing with few exemplars. The editing exemplar, containing the original and modified support code snippets, showcases a certain editorial style and implies the edit intention for a query code snippet. To achieve this, we propose a machine learning approach to adapt the editorial style derived from few exemplars to a query code snippet. Our learning approach combines edit representations extracted from editing exemplars and compositionally generalizes them to the query code snippet editing via multi-extent similarities ensemble. Specifically, we parse the code snippets using language-specific grammar into abstract syntax trees. We apply the similarities measurement in multiple extents from individual nodes to collective tree representations, and ensemble them through a similarity-ranking error estimator. We evaluate the proposed method on two datasets in C\# and Python languages and respectively show 8.0\% and 10.9\% absolute accuracy improvements compared to baselines.
PDF Abstract