Retrieval-Based Neural Code Generation

Abstract

In models to generate program source code from natural language, representing this code in a tree structure has been a common approach. However, existing methods often fail to generate complex code correctly due to a lack of ability to memorize large and complex structures. We introduce ReCode, a method based on subtree retrieval that makes it possible to explicitly reference existing code examples within a neural code generation model. First, we retrieve sentences that are similar to input sentences using a dynamic-programming-based sentence similarity scoring method. Next, we extract n-grams of action sequences that build the associated abstract syntax tree. Finally, we increase the probability of actions that cause the retrieved n-gram action subtree to be in the predicted code. We show that our approach improves the performance on two code generation tasks by up to +2.6 BLEU.

arXiv commentThis paper is accepted in EMNLP 2018. It has 6 pages

References in corpus (5)

• Incorporating Copying Mechanism in Sequence-to-Sequence Learning
• Language to Logical Form with Neural Attention
• A Syntactic Neural Model for General-Purpose Code Generation
• Latent Predictor Networks for Code Generation
• Guiding Neural Machine Translation with Retrieved Translation Pieces

Cited by in corpus (6)

• Unifying Human and Statistical Evaluation for Natural Language Generation
• Relevance Transformer: Generating Concise Code Snippets with Relevance Feedback
• TreeGen: A Tree-Based Transformer Architecture for Code Generation
• One-Shot Learning for Text-to-SQL Generation
• Semantic Scaffolds for Pseudocode-to-Code Generation
• Question Answering via Web Extracted Tables and Pipelined Models