Siamese-Based BiLSTM Network for Scratch Source Code Similarity Measuring

As a popular block-based programming language, Scratch attracts considerable attention in society and educational fields. Code similarity measuring is a major research direction in Scratch, which plays a significant role in clone detection and project recommendation. However, there are few studies focusing on it. In this paper, we propose a Siamese-Based bidirectional Long Short-Term Memory (BiLSTM) network to solve this problem. Specifically, a token-based code representation scheme is designed to abstract the blocks in Scratch. Then the obtained token stream is fed to a word embedding model for training. Next, we devise an improved Siamese-based BiLSTM model to measure the source code similarity. Finally, in order to evaluate the performance of proposed model, we construct a dataset from Scratch official website. The results show that it achieves more than 90% accuracy and recall. In addition, the proposed model is applied in the code cluster task, and reaches 95% accuracy.