Structure-Preserving Document-Level Data Augmentation via Discourse Parsing
Data augmentation for document-level NLP risks breaking inter-sentential coherence and invalidating sentence-aligned labels. This paper proposes an LLM-driven discourse parsing framework that preserves sentence order, count, and rhetorical relations $R_{i,j}$ while replacing key phrases $\mathbf{K}_i$ to generate augmented documents $\mathbf{D}'$. By keeping the $n$-sentence structure fixed, labels such as summary indices or QA spans can be transferred without expensive LLM relabeling.
The paper offers a conceptually clean solution to a real problem: document-level DA often corrupts discourse structure, which invalidates supervision. The authors empirically show that enforcing rhetorical constraints and fixed sentence alignment improves combined-data training across summarization and QA. However, the evidence is undermined by small author-annotated validation sets, the absence of statistical significance tests, and an entirely prompt-based pipeline whose internals are undisclosed. The work is a promising proof of concept, but readers are asked to trust a black-box LLM workflow.
The structural-preservation argument is well-motivated and empirically supported. Comparisons with Direct LLM Augmentation and LLM-based relabeling show that unconstrained generation consistently introduces supervision noise, whereas the proposed constrained pipeline limits this drift. The ablation study also confirms that discourse labels matter: removing relations degrades performance on every tested dataset. Finally, cross-architecture experiments with BertSum, SimCLS, RoBERTa, and ELECTRA demonstrate that the gains are not an artifact of the Qwen fine-tuning setup.
Reproducibility and evaluation rigor are the largest weaknesses. The core functions $\alpha$, $\beta$, $\gamma$, $\delta$ are realized through undisclosed LLM prompts: the authors state that 'The augmentation process is decomposed into three independently prompted stages, each executed with a single prompt,' yet no prompt text is provided. Discourse parsing quality is validated on only 100 author-annotated sentence pairs per language, a sample far too small to support robust claims. No statistical significance tests are reported, so gains such as BLEU moving from $0.269$ to $0.276$ on the Chinese corpus may be noise. The paper also silently ignores that on SQuAD the Inp+Con baseline drops in F1 from $0.780$ (Inp) to $0.746$, suggesting the control set is not always beneficial.
The evidence supports the high-level claim that structure preservation helps, but several comparisons are tilted in the authors' favor. The UniRST baseline is handicapped by a hand-engineered mapping that collapses its fine-grained relations into the authors' six categories; they admit that this mapping is 'designed specifically for our augmentation framework and is not intended as an official equivalence between the original discourse inventories.' Meanwhile, Back-translation remains surprisingly competitive: on Newsroom it achieves BLEU $0.752$ under Aug, outperforming the proposed method's $0.639$, and in combined settings it still rivals the proposed approach. The gains over Direct LLM Augmentation do support the value of discourse constraints, yet the margin varies widely across datasets and metrics, undermining claims of uniform robustness.
Independent reproduction is severely obstructed by missing artifacts. While downstream training hyperparameters are listed in detail (LoRA rank $8$, learning rate $1 \times 10^{-5}$), no generation hyperparameters are given for the augmentation LLM (temperature, top-$p$, sampling strategy), and the prompt templates for discourse parsing, key-phrase replacement, and document generation are omitted. No code or data release is mentioned. Moreover, the experiments rely on custom $2{,}000/2{,}000/1{,}000$ input/control/test splits rather than standard splits, so direct comparison with prior published numbers is unreliable. Reproducing the pipeline would require substantial guesswork.
Data augmentation (DA) plays a vital role in improving model performance in low-resource or limited-supervision scenarios within Natural Language Processing (NLP). Existing DA methods, such as synonym replacement and back-translation, have demonstrated effectiveness on the word or sentence level; however, they frequently neglect discourse-level coherence and logical flow, which are essential for complex tasks dependent on inter-sentential relationships. In this paper, we propose a structure-preserving document-level data augmentation framework based on large language models (LLMs) and document-level discourse structure parsing. Our approach identifies rhetorical relations between sentence pairs and extracts key phrases, which are then replaced with semantically coherent but topically shifted content while preserving the original discourse structure. Experimental results on text summarization and question answering show that the proposed augmented data provides complementary supervision and improves combined-data training across the evaluated tasks. Comparisons with a standard RST parser and controlled ablations further support the proposed discourse representation.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.