From Execution to Embedding: Enriching Code Representations with Data Difference Signals for Comment Generation

Modern automatic comment generation tools often fail to capture data-centric workflow intents, because syntax alone provides weak signals of how code transforms data. We instead capture semantic data differences, symbolic descriptions of post-execution data transformations, and integrate them with code through a dual-encoder architecture. We evaluate this approach on a dataset of executed Python notebooks pairing code, effect sequences, and human comments. When no transformation is detected, the full pipeline outperforms the baseline across both automatic metrics and human evaluation. When transformations are present, the baseline remains competitive, though some simplified variants surpass it on specific metrics. We make available all software and data to encourage replication and further studies in this area. While our experiments focus on comment generation, our core contribution is broader: we introduce execution-aware embeddings and argue for their applicability to a variety of downstream tasks.