SkCoder: A Sketch-based Approach for Automatic Code Generation

TL;DR

SkCoder addresses automatic code generation by emulating developer code reuse through a sketch-based pipeline: retrieve a similar snippet using BM25, extract a relevant code sketch from the retrieved code, and edit the sketch into the final code guided by the NL requirement. The approach disentangles content selection from final assembly via three modules—Retriever, Sketcher, and Editor—and uses a two-stage training regime due to non-differentiable sketching. Empirical results on HearthStone, Magic, and AixBench-L show SkCoder outperforming state-of-the-art baselines (e.g., EM gains up to 22.41% and Pass@1 gains up to 35.39% on certain datasets), with ablations confirming the value of each component and the sketch design. Human evaluation corroborates improvements in correctness, code quality, and maintainability, underscoring the practical impact of code sketches for more reliable NL-to-code generation.

Abstract

Recently, deep learning techniques have shown great success in automatic code generation. Inspired by the code reuse, some researchers propose copy-based approaches that can copy the content from similar code snippets to obtain better performance. Practically, human developers recognize the content in the similar code that is relevant to their needs, which can be viewed as a code sketch. The sketch is further edited to the desired code. However, existing copy-based approaches ignore the code sketches and tend to repeat the similar code without necessary modifications, which leads to generating wrong results. In this paper, we propose a sketch-based code generation approach named SkCoder to mimic developers' code reuse behavior. Given a natural language requirement, SkCoder retrieves a similar code snippet, extracts relevant parts as a code sketch, and edits the sketch into the desired code. Our motivations are that the extracted sketch provides a well-formed pattern for telling models "how to write". The post-editing further adds requirement-specific details to the sketch and outputs the complete code. We conduct experiments on two public datasets and a new dataset collected by this work. We compare our approach to 20 baselines using 5 widely used metrics. Experimental results show that (1) SkCoder can generate more correct programs, and outperforms the state-of-the-art - CodeT5-base by 30.30%, 35.39%, and 29.62% on three datasets. (2) Our approach is effective to multiple code generation models and improves them by up to 120.1% in Pass@1. (3) We investigate three plausible code sketches and discuss the importance of sketches. (4) We manually evaluate the generated code and prove the superiority of our SkCoder in three aspects.

Figures (6)

• Figure 1: The process of reusing the similar code by developers.
• Figure 2: The illustration of how developers reuse the similar code. The relevant content in the similar code is highlighted.
• Figure 3: The overview of our approach.
• Figure 4: An illustration of our sketch.
• Figure 5: Examples of three sketches.