What You See is What it Means! Semantic Representation Learning of Code based on Visualization

in ACM Transactions on Software Engineering and Methodology (2021)

Recent successes in training word embeddings for NLP tasks have encouraged a wave of research on representation learning for sourcecode, which builds on similar NLP methods. The overall objective is then ... [more ▼]

Recent successes in training word embeddings for NLP tasks have encouraged a wave of research on representation learning for sourcecode, which builds on similar NLP methods. The overall objective is then to produce code embeddings that capture the maximumof program semantics. State-of-the-art approaches invariably rely on a syntactic representation (i.e., raw lexical tokens, abstractsyntax trees, or intermediate representation tokens) to generate embeddings, which are criticized in the literature as non-robustor non-generalizable. In this work, we investigate a novel embedding approach based on the intuition that source code has visualpatterns of semantics. We further use these patterns to address the outstanding challenge of identifying semantic code clones. Wepropose theWySiWiM(“What You See Is What It Means”) approach where visual representations of source code are fed into powerfulpre-trained image classification neural networks from the field of computer vision to benefit from the practical advantages of transferlearning. We evaluate the proposed embedding approach on the task of vulnerable code prediction in source code and on two variationsof the task of semantic code clone identification: code clone detection (a binary classification problem), and code classification (amulti-classification problem). We show with experiments on the BigCloneBench (Java), Open Judge (C) that although simple, ourWySiWiMapproach performs as effectively as state of the art approaches such as ASTNN or TBCNN. We also showed with datafrom NVD and SARD thatWySiWiMrepresentation can be used to learn a vulnerable code detector with reasonable performance(accuracy∼90%). We further explore the influence of different steps in our approach, such as the choice of visual representations or theclassification algorithm, to eventually discuss the promises and limitations of this research direction. [less ▲]