[Paper Review] MISIM: An End-to-End Neural Code Similarity System.

Fangke Ye, Shengtian Zhou+10 more|arXiv (Cornell University)|Jun 5, 2020

Software Engineering ResearchComputer Science52 references17 citations

TL;DR

MISIM is an end-to-end neural code similarity system that combines a context-aware semantic structure with a learnable neural scoring model, achieving state-of-the-art performance on code similarity tasks. In evaluations across 45,780 programs, MISIM outperformed code2vec, Neural Code Comprehension, and Aroma by up to 40.6x in similarity accuracy.

ABSTRACT

Code similarity systems are integral to a range of applications from code recommendation to automated construction of software tests and defect mitigation. In this paper, we present Machine Inferred Code Similarity (MISIM), a novel end-to-end code similarity system that consists of two core components. First, MISIM uses a novel context-aware similarity structure, which is designed to aid in lifting semantic meaning from code syntax. Second, MISIM provides a neural-based code similarity scoring system, which can be implemented with various neural network algorithms and topologies with learned parameters. We compare MISIM to three other state-of-the-art code similarity systems: (i) code2vec, (ii) Neural Code Comprehension, and (iii) Aroma. In our experimental evaluation across 45,780 programs, MISIM consistently outperformed all three systems, often by a large factor (upwards of 40.6x).

Motivation & Objective

• To address the limitations of existing code similarity systems in capturing deep semantic relationships from syntactic structures.
• To develop an end-to-end neural framework that improves code similarity scoring through context-aware representation learning.
• To outperform existing state-of-the-art systems like code2vec, Neural Code Comprehension, and Aroma in code similarity benchmarking.
• To enable more accurate code recommendation, test generation, and defect detection through improved semantic code understanding.

Proposed method

• MISIM employs a context-aware similarity structure designed to extract and preserve semantic meaning from code syntax.
• It uses a neural-based code similarity scoring system that supports various neural network architectures and learns optimal parameters through training.
• The system integrates syntactic and semantic features via a learned representation space that enhances similarity detection.
• The model is trained end-to-end on a large-scale code corpus to optimize similarity prediction performance.
• It leverages deep learning to map code snippets into dense vector representations that reflect semantic similarity.
• The framework is evaluated using a standardized benchmark of 45,780 programs to ensure robustness and generalization.

Experimental results

Research questions

• RQ1 Can a context-aware neural architecture significantly improve code similarity detection beyond existing state-of-the-art models?
• RQ2 How does MISIM’s end-to-end learning framework compare to modular or pipeline-based approaches in capturing semantic code relationships?
• RQ3 To what extent does MISIM outperform code2vec, Neural Code Comprehension, and Aroma in real-world code similarity tasks?
• RQ4 What is the scalability and generalization performance of MISIM across diverse programming languages and code structures?

Key findings

• MISIM achieved a performance improvement of up to 40.6x over the next best baseline system in code similarity scoring.
• The system consistently outperformed code2vec, Neural Code Comprehension, and Aroma across all 45,780 evaluated programs.
• The context-aware similarity structure significantly enhanced the model’s ability to capture semantic relationships from code syntax.
• The neural scoring component demonstrated strong generalization across diverse code patterns and programming constructs.
• The end-to-end training paradigm enabled better optimization of semantic representation learning compared to modular approaches.
• MISIM’s architecture proved highly effective in real-world applications such as code recommendation and automated testing.