An alternative to code comment generation? Generating comment from bytecode

Context: Due to the importance and necessity of code comments, recent works propose many comment generation models with source code as input. But sometimes there has no access to obtain the source code, only the bytecode, such as many Apps. Objective: If there is away to generate comments for bytecode directly, tasks such as malware detection and understanding closed-source software can benefit from the generated comment because it improves the understandability of the system. Therefore, we propose a novel approach called ByteGen to generate comments from bytecode. Methods: Specifically, to extract the structure characteristic of the bytecode, we utilize the control flow graph (CFG) of the bytecode and use a special traversal named enhanced SBT to serialize CFG. The enhanced SBT can completely preserve the structure of the CFG in a sequence. We set up experiments on a dataset with a scale of about 50,000 bytecode-comment pairs collected from Maven. Results: Experimental results show that the average BLEU-4 score of ByteGen is 28.67, which outperforms several baselines, and a human study also indicates the effectiveness of ByteGen in generating comments from bytecodes. Conclusion: In general, ByteGen performs better than other baselines. Therefore, this also proves the effectiveness of our approach in the code comment generation scenario without source code.