Lost in Translation, Found in Summary: Summary-Driven Supervision for Cross-Language Code Retrieval

Abstract

Cross-language code retrieval (CLCR) is critical to modern software development, and recent studies have shown that contrastive learning is an effective way to learn cross-language code semantics. However, these methods are still limited by two fundamental challenges. First, they rely on online judge platforms or multilingual NL-to-code datasets, which either lack diversity or fail to ensure semantic alignment between code snippets, thereby limiting the effectiveness of cross-language semantic learning. Second, during training, code snippets from different languages are often encoded into heterogeneous representation distributions. This distribution inconsistency introduces similarity bias and weakens the model's ability to capture accurate cross-language semantics. To address these challenges, we propose SemSIR, an end-to-end framework for CLCR. The key insight is to use Summary IR, a natural-language intermediate representation, to bridge semantic gaps across languages. Built on this insight, SemSIR automatically constructs scalable cross-language supervision from real-world repositories, and further combines distribution alignment with CLIP-style contrastive training to learn more discriminative language-independent code semantics. In our evaluation, we collect 24 million real-world code snippets across 9 languages and automatically construct 7 million semantic groups. Manual inspection of 1,000 sampled groups shows that 92.7% are fully semantically coherent, with an average Group Purity of 96.8%. On CLCR, SemSIR consistently outperforms all baselines on two benchmarks, surpassing the state-of-the-art method by up to 24.7% in MRR and 50.1% in MAP on average. In addition, experimental results on a security case study of tracing rewritten malware across languages show that SemSIR achieves 92.20 MRR and 88.40 MAP, highlighting its potential for security-oriented CLCR.