PaperlayerView exact revision on arXiv

Source-linked AI summary

Diversity-oriented Data Augmentation with Large Language Models

Zaitian Wang, Jinghan Zhang, Xinhao Zhang, Kunpeng Liu, Pengfei Wang, Yuanchun Zhou

arXiv:2502.11671v2cs.CLcs.AIcs.LG

TL;DR

Many generative augmentation methods emphasize adding samples while giving limited attention to dataset diversity, despite diversity's place among the stated qualities of high-quality data. DoAug fine-tunes an LLM paraphraser for diverse, high-affinity outputs, and experiments across 12 datasets report improved learning performance while increasing diversity and maintaining coherence.

  • Problem

    Many generative data augmentation methods focus on enlarging datasets while giving limited consideration to how augmentation affects dataset diversity.

    Source: §1 · p. 2

  • Method

    DoAug fine-tunes an LLM paraphraser using preferences for diverse outputs and DPO, then applies it to selected important samples.

    Source: §1 · p. 2

  • Results

    Across 12 datasets, DoAug significantly benefits learning performance while increasing dataset diversity and maintaining coherence.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The authors report that DoAug increases diversity across six measurements while maintaining high affinity between augmented and original data.

    Source: §1 · p. 2

    The authors identify demographic bias and plausible but factually incorrect generations as potential risks of LLM-based augmentation.

    Source: §7 · p. 10

Abstract

from arXiv · show

Data augmentation is an essential technique in natural language processing (NLP) for enriching training datasets by generating diverse samples. This process is crucial for improving the robustness and generalization capabilities of NLP models. However, a significant challenge remains: \textit{Insufficient Attention to Sample Distribution Diversity}. Most existing methods focus on increasing the sample numbers while neglecting the sample distribution diversity, which can lead to model overfitting. In response, we explore data augmentation's impact on dataset diversity and propose a \textbf{\underline{D}}iversity-\textbf{\underline{o}}riented data \textbf{\underline{Aug}}mentation framework (\textbf{DoAug}). % \(\mathscr{DoAug}\) Specifically, we utilize a diversity-oriented fine-tuning approach to train an LLM as a diverse paraphraser, which is capable of augmenting textual datasets by generating diversified paraphrases. Then, we apply the LLM paraphraser to a selected coreset of highly informative samples and integrate the paraphrases with the original data to create a more diverse augmented dataset. Finally, we conduct extensive experiments on 12 real-world textual datasets. The results show that our fine-tuned LLM augmenter improves diversity while preserving label consistency, thereby enhancing the robustness and performance of downstream tasks. Specifically, it achieves an average performance gain of \(10.52\%\), surpassing the runner-up baseline with more than three percentage points.

1 Introduction

Existing augmentation methods often enlarge datasets without adequately addressing diversity, while diverse, coherent training data is a stated quality goal. DoAug fine-tunes an LLM paraphraser to generate diverse, high-affinity samples and improves learning performance across 12 datasets.

1 Introduction

  • The introduction characterizes high-quality datasets as large, coherent, and diverse, with linguistic variation supporting robustness across real-world conditions.Examples include variation in dialect, tone, formality, and domain-specific terms.Source: §1 · pp. 1–2
  • Many existing generative augmentation methods prioritize dataset size and give limited attention to diversity, while simple perturbations can introduce noise or redundancy.For example, deleting “not” can compromise label integrity.Source: §1 · p. 2
  • DoAug uses a fine-tuned LLM to paraphrase sentences while preserving essential meaning, aiming to maintain affinity between original and augmented samples.The paper presents this affinity as a way to minimize augmentation's influence on dataset coherence.Source: §1 · p. 2
  • DoAug fine-tunes an LLM paraphraser on preferences for more diverse rather than repetitive outputs and applies it to a selected coreset of important samples.The framework uses DPO to encourage diverse generation while limiting augmentation to selected samples to reduce computational overhead and cost.Source: §1 · p. 2
  • Across 12 datasets, DoAug significantly benefits learning performance while increasing dataset diversity and maintaining coherence.The authors also report higher diversity on six measurements, high affinity to original data, and significant downstream performance improvements.Source: §1 · p. 2

2 Related Work

Related work reviews textual augmentation methods ranging from perturbations to LLM rewriting, then surveys diversity evaluation and the limited integration of diversity into coherence-oriented LLM augmentation.

2.1 Textual Data Augmentation

  • Textual augmentation ranges from character or word perturbations to syntax changes, back-translation, masked-word substitution, and LLM-based rewriting.The review describes AugGPT and Self-LLMDA as LLM approaches for rewriting or selecting prompts for augmented samples.Source: §2.1 · pp. 2–3
  • Prior methods such as AugGPT preserve coherence, but the reviewed approaches generally neglect dataset diversity; prompt-based diversity incentives are also discussed.The review cites three types of prompt-based incentives for diversity.Source: §2.1 · p. 3

2.2 Dataset Diversity Evaluation

  • Dataset diversity has no unified definition, and evaluation approaches include token-level, embedding-level, and human measures.The review also describes embedding-space metrics and findings linking diversity and affinity to model performance and generalization.Source: §2.2 · p. 3
  • Although diversity has informed some augmentation methods, the review identifies its lack of integration with coherence-ensured LLM augmentation such as AugGPT.Prior work also investigates diversity and affinity jointly in augmentation.Source: §2.2 · p. 3

3 Methodology

DoAug formulates augmentation around maximizing dataset-diversity gain while seeking performance gain under fixed conditions. It fine-tunes an LLM to produce diverse, meaning-preserving paraphrases, selects important samples for augmentation, and retains the most diverse generated paraphrases.

3.1 Problem Formulation

  • DoAug defines diversity gain as the change from original to augmented dataset diversity and aims to maximize it; under fixed conditions, it expects the diversity-optimal augmenter to achieve maximum performance gain.Performance gain is defined as augmented-dataset performance minus original-dataset performance.Source: §3.1 · p. 3

3.2 Framework Overview

  • The framework fine-tunes an LLM as a paraphraser, selects important samples as augmentation seeds, and integrates diverse paraphrases with original samples.The LLM is first instruction fine-tuned to preserve sentence semantics, then trained on a preference dataset with DPO to encourage diverse generation.Source: §3.2 · pp. 3–4

3.3 Diverse Paraphraser Fine-tuning

  • The LLM paraphraser rewrites sentences with alternative expressions while maintaining semantics, aiming to increase linguistic diversity while preserving augmented data-label coherence.The paper attributes diversity to changed expressions and coherence to the LLM capturing sentence semantics and being prompted to paraphrase.Source: §3.3 · p. 4

3.3.1 LLM Paraphraser Training with PEFT

  • Supervised instruction fine-tuning trains the LLM to paraphrase existing sentences, while LoRA freezes pretrained weights and updates low-rank matrices to reduce training cost.The SFT data are sampled from the ChatGPT Paraphrases dataset.Source: §3.3.1 · p. 4

3.3.2 LLM Generation Diversity Enhancement with DPO

  • For DPO preferences, the most distant of five embedded paraphrases is chosen and the most similar rejected; DPO trains toward preferred outputs without a separate reward model.DPO aims to increase the probability of preferred outputs and decrease that of dispreferred outputs.Source: §3.3.2 · pp. 4–5

3.3.3 Diversity-oriented Sampling

  • For each input, beam search generates K sequences, ranks them by distance from the source, and retains the most distant to reduce redundancy and prevent overfitting.The algorithm first trains the LLM with SFT, then fine-tunes it on the DPO preference dataset.Source: §3.3.3 · p. 5

3.4 Selective Coreset Data Augmentation

  • DoAug scores samples using training dynamics and post-training metrics, then hierarchically prunes low-importance samples, retains middle-importance samples, and augments high-importance ones.Only high-importance samples seed augmentation; their originals and paraphrases are combined with the retained middle-importance samples.Source: §3.4 · p. 5

4 Experiments

Across 12 text-classification datasets, DoAug combines strong downstream performance with increased dataset diversity, while evaluating affinity, validity, component ablations, and alternative model architectures.

4.1.1 Evaluation Criterion

  • Evaluation measures latent diversity at sample and dataset levels, lexical diversity, embedding-deviation affinity, and downstream classification accuracy.Diversity measures include pairwise embedding distance and cosine similarity, embedding coverage and uniformity, vocabulary size, and unique 3-grams.Source: §4.1.1 · p. 6

4.1.2 Datasets; 4.1.3 Baseline Methods; 4.1.4 Key Implementation Information

  • Experiments evaluate DoAug on 12 NLP datasets covering varied text-classification tasks, sampling 1.2K examples from each full dataset to unify evaluation and simulate low-resource conditions, and compare it with twelve representative methods.For each original sentence, five paraphrases are generated and the most diversified output is sampled; the coreset ratio raugment : rretain : rprune is 1 : 1 : 1.Source: §4.1.2 · p. 6, §4.1.3 · p. 6, §4.1.4 · p. 6

4.2 Overall Results

  • DoAug ranks highest in downstream performance and diversity and places at the top in combined diversity and affinity rankings.Figure 3 averages results across 12 datasets and diversity rankings across six metrics; lower rankings are better.Source: §4.2 · p. 7, Figure 3 · p. 7

4.3.1 Performance Gains

  • 10.52% average performance gain, outperforming the runner-up by 3.76 percentage points and beating baselines on 11 of 12 datasets; Table 1 reports accuracy results.RTE is the exception.Source: §4.3.1 · p. 7, Table 1 · p. 7

4.3.2 Diversity Gain

  • DoAug scores 0.98 on average across six diversity metrics, leading on Distance, Dispersion, Vocabulary Size, and Unique 3-grams.It is competitive on Isocontour Radius and Homogeneity; Chain, Hint, and Taboo also achieve reasonably good diversity gains.Source: §4.3.2 · p. 7

4.3.3 Affinity and Paraphrase Validity

  • DoAug’s affinity exceeds the baselines except Unmask, whose affinity score is 1.95; 95% of paraphrases were human-judged valid and 97% were LLM-judged valid.The authors attribute Unmask’s exceptionally high affinity to its BERT-based recovery and evaluation using BERT embeddings.Source: §4.3.3 · p. 8, Figure 4 · p. 8

4.4 Ablation Studies

  • Ablations show every framework component contributes to performance, with coreset-focused augmentation the most important; diversity sampling affects sample-level latent diversity more, while DPO affects lexical diversity more.Replacing DPO with higher-temperature sampling or diversity-incentive prompts yields inferior results.Source: §4.4 · p. 8, Figure 5 · p. 8

4.5 LLM Architectures Adaptability

  • With Qwen2.5-1.5B-Instruct as augmenter, DoAug improves diversity over original data and achieves accuracy comparable to Llama; it also benefits GPT-2 and T5-large classifiers.The classifier tests use MNLI, CoLA, and RCT.Source: §4.5 · p. 9, Figure 7 · p. 9, Table 4 · p. 9

5 Conclusion

DoAug treats diversity as an often-neglected quality of textual datasets and uses a fine-tuned LLM paraphraser to increase it while preserving coherence. Experiments report that the diversity gains benefit downstream learning performance.

5 Conclusion

  • DoAug fine-tunes an LLM paraphraser to rewrite seed sentences into high-affinity samples, then uses DPO preference training to encourage diverse generation.The preference dataset favors more diverse paraphrases over repetitive ones.Source: §5 · p. 9
  • The framework targets dataset diversity, which the authors describe as an important but often-ignored factor in high-quality datasets.DoAug aims to enlarge textual datasets while maintaining coherence.Source: §5 · p. 9
  • Experiments report that DoAug’s diversity gains significantly benefit downstream task learning performance.The passage does not provide a numeric performance comparison.Source: §5 · p. 9

7 Limitations

The authors identify limitations in diversity and augmentation evaluation, generation settings, benchmark scope, and LLM-related risks. Their evaluation criteria span embedding-level and lexical diversity measures, alongside an affinity measure.

7 Limitations

  • The diversity assessment uses only a subset of metrics and omits human-centered evaluation, so it may not capture every aspect of diversity.The authors note that universally accepted diversity metrics are lacking.Source: §7 · p. 9
  • Human and task-aware LLM evaluations have constraints, including potential biases and difficulty capturing nuanced correctness in some contexts.The authors state that neither human nor LLM evaluation is perfect.Source: §7 · p. 10
  • Only two settings were tested for each generation factor, leaving temperature, prompts, and LLM choice incompletely explored.The authors say this restriction may have affected generated-sample diversity and quality.Source: §7 · p. 10
  • Evaluation focused mainly on sentence classification, included two QA-based reasoning tasks, and used mostly English datasets; multimodal scenarios were not explored.The authors also identify mathematical reasoning, instruction following, creative writing, and chain-of-thought reasoning as untested task areas.Source: §7 · p. 10
  • LLM augmentation may introduce demographic bias or factual inaccuracies, which can degrade task-model reliability and accuracy, especially in healthcare or finance.The authors call for rigorous validation, bias-detection frameworks, and human oversight.Source: §7 · p. 10
  • The authors identify more comprehensive diversity metrics and evaluation across diverse modalities as future-work priorities.Source: §7 · p. 10

A Evaluation Criterion

  • Embedding-level diversity criteria include sample distance and dispersion, cluster spread via isocontour radius, and distribution uniformity via homogeneity.Distance and dispersion are described so that larger values indicate greater diversity; homogeneity reflects uniformity of the cluster distribution.Source: A Evaluation Criterion · p. 14
  • Lexical diversity is measured with vocabulary size and unique 3-grams, while affinity reflects deviations between augmented and original class centers.Invalid words are excluded from lexical diversity calculations.Source: A Evaluation Criterion · p. 14

B Dataset Specification

The experiments use 12 textual datasets spanning entailment, medical, sentiment, acceptability, equivalence, and subjectivity tasks. Table 5 summarizes their domains, tasks, input schemes, and class counts.

B Dataset Specification

  • The study evaluates 12 NLP datasets covering textual entailment, medical classification, sentiment, acceptability, sentence equivalence, and subjectivity tasks.The medical datasets include ChemProt, RCT, and Symptoms; sentiment datasets include MPQA, SST-2, and Yelp.Source: B Dataset Specification · p. 14
  • Table 5 summarizes the 12 datasets by domain, application task, input scheme, and number of classes.Source: B Dataset Specification · p. 14, Table 5 · p. 15
  • The dataset set includes GLUE tasks alongside ANLI, ChemProt, RCT, MPQA, SUBJ, Yelp, and Symptoms.The passage identifies Hugging Face as a source for the GLUE datasets and several other datasets, and Kaggle for Symptoms.Source: B Dataset Specification · p. 14

C Implementation Details

The implementation uses a Llama-3.2-1B-Instruct paraphraser trained with supervised fine-tuning and DPO, with importance-based coreset selection and five generated candidates per sentence. Baselines use matched coreset ratios and final sample counts; examples show added details and vocabulary.

C Implementation Details

  • The Llama-3.2-1B-Instruct paraphraser is trained with SFT and DPO, using LoRA adapters and 100,000 SFT sentence pairs plus 50,000 DPO preference pairs.The authors report 32 minutes for SFT and 36 minutes for DPO training; the augmenter paraphrases roughly one sentence per second.Source: C Implementation Details · p. 16
  • For each sentence, the implementation generates five paraphrases and samples the most diversified output, with coreset ratios of 1:1:1 for augmentation, retention, and pruning.The downstream BERT model is trained for three epochs, and experiments are repeated with ten random seeds.Source: C Implementation Details · p. 16

D Baseline Methods Implementation

  • Baselines use the same coreset ratios and final sample counts as DoAug, but do not rank samples before selection; the original data are not augmented.The authors use matching random seeds to support fair comparisons.Source: D Baseline Methods Implementation · p. 16

E Augmentation Examples

  • Augmentation examples show DoAug adding details such as “have trouble moving” and vocabulary such as “fondness” and “sophisticated.”The examples are from Symptoms and SST-2.Source: E Augmentation Examples · p. 16

F Diversity Evaluation

The section directs readers to Table 7 for the original diversity scores of all methods, averaged across 12 datasets.

F Diversity Evaluation

  • Table 7 reports six diversity metrics averaged across 12 datasets, and the section identifies it as the source of the methods’ original diversity scores.The supplied passages do not state the metric values.Source: Table 7 · p. 16, F Diversity Evaluation · p. 16

G Measurements for Preference Dataset Construction and Diversity Sampling

The authors compare Euclidean distance with cosine similarity for constructing preference data and sampling diverse paraphrases. They report high agreement between the metrics and use Euclidean distance in their final settings.

G Measurements for Preference Dataset Construction and Diversity Sampling

  • 97.1% of the time, Euclidean distance and cosine similarity select the same most-diverse chosen samples; for most-repetitive rejected samples, they agree in 97.8% of cases.The comparison used a subset of 1,000 original paraphrase-dataset samples.Source: G Measurements for Preference Dataset… · p. 17
  • 0.96 is the Pearson correlation between Euclidean distance and cosine dissimilarity across 5,000 paraphrases, each containing five paraphrases.The comparison measures distance or dissimilarity from the original sentence.Source: G Measurements for Preference Dataset… · p. 17
  • Euclidean distance was selected for the final settings because it is more straightforward to interpret, with greater distance indicating greater diversity.The authors expect performance to remain consistent if cosine similarity is used instead.Source: G Measurements for Preference Dataset… · p. 17
  • Figure 8a plots 500 sampled points, while Figure 8b normalizes all scores to [0, 1].The distance relationship is shown in Figure 8a; Figure 8b shows the score distribution.Source: Figure 8 · p. 17, G Measurements for Preference Dataset… · p. 17

H Full Results of Replacing DPO training

The section reports parameter and coreset sensitivity findings, alongside comparisons with larger training subsets. It also points to a six-metric diversity evaluation of replacing DPO training.

H Full Results of Replacing DPO training

I Parameter Sensitivity Study

  • Because LLM training is costly, the authors follow existing or library-default settings and test sensitivity for parameters unique to DoAug.The section introduces the parameter sensitivity studies.Source: I Parameter Sensitivity Study · p. 17

I.1 Number of Generated Sentences

  • The best performance occurs when K, the number of generated sentences per original, is between 5 and 8.Too-small K limits generation diversity, while too-large K can move paraphrases too far from the original semantics and harm label preservability.Source: §I.1 · p. 17

I.2 Coreset Ratio

  • A pruning ratio of 1/3 and augmentation ratio of 1/2 yield the best performance, producing equal one-third shares of augmented, retained, and pruned samples.The ratios are measured as described in Figure 10: pruning relative to pruned, retained, and augmented samples, and augmentation relative to retained and augmented samples.Source: §I.2 · p. 17, Figure 10 · p. 18

J Coreset Selection Methods

  • “Variance” and “CCS w/ AUM” benefit the most datasets, helping 9 of 12; even suboptimal coresets usually outperform augmentation without coreset targeting.Different datasets favor different coreset methods, and performance usually does not significantly degenerate with suboptimal choices.Source: J Coreset Selection Methods · p. 18, Figure 11 · p. 18

K Alleviating the Low-resource Problem

  • DoAug with 800 original and 400 augmented samples alleviates the low-resource problem and sometimes performs better than 1.2K-sample subsets.Table 8 compares low-resource datasets (800), larger original subsets (1.2 K), and DoAug (800 Original + 400 Augmented Samples).Source: K Alleviating the Low-resource Problem · p. 18, Table 8 · p. 18

L Full Results of LLM Architecture Exploration

The section points to full results for Qwen performance on 12 datasets and diversity comparisons between Llama and Qwen. Additional tests report benefits on reasoning and multilingual tasks.

L Full Results of LLM Architecture Exploration

M Applicability on More Datasets

  • DoAug also benefits two multiple-choice reasoning datasets and a multilingual sentiment dataset tested in English, Spanish, German, French, Italian, and Portuguese.Arabic samples were removed because the Llama augmenter does not support Arabic; Code-Mixed was also included for multilingual sentiment analysis.Source: M Applicability on More Datasets · p. 19, Figure 14 · p. 19
Loading 2502.11671v2…