Traditional generative models are limited to predicting sequences of terminal tokens. However, ambiguities in the generation task may lead to incorrect outputs. Towards addressing this, we introduce Grammformers, transformer-based grammar-guided models that learn (without explicit supervision) to generate sketches -- sequences of tokens with holes. Through reinforcement learning, Grammformers learn to introduce holes avoiding the generation of incorrect tokens where there is ambiguity in the target task. We train Grammformers for statement-level source code completion, i.e., the generation of code snippets given an ambiguous user intent, such as a partial code context. We evaluate Grammformers on code completion for C# and Python and show that it generates 10-50% more accurate sketches compared to traditional generative models and 37-50% longer sketches compared to sketch-generating baselines trained with similar techniques.
Download