Data augmentation (DA) is a common technique for improving model performance when labeled data is scarce. In NLP, methods such as synonym replacement are widely used, but they often operate at the word or sentence level and can break the overall coherence of a document. A new arXiv preprint (2610.19497) addresses this limitation by introducing a structure-preserving DA method based on discourse parsing.

The paper argues that by parsing the discourse structure of a document, it is possible to generate augmented versions that maintain the original rhetorical relationships between sentences and paragraphs. This stands in contrast to simpler augmentation techniques that ignore document-level organization. The approach is aimed at low-resource or limited-supervision settings, where preserving structure could help models learn more effectively from augmented data.

As the source is an abstract only, specific experimental results and implementation details are not yet available. The authors position their work as a response to the shortcomings of existing methods, suggesting that discourse-aware augmentation may be a promising direction for document-level NLP tasks.