LLM text augmentation falls short of classical methods in imbalanced classification benchmarks

A new benchmark study has compared various text augmentation methods for imbalanced text classification, evaluating both classical techniques like EmbSMOTE and newer LLM-based approaches. The research found that LLM-generated data was statistically equivalent or inferior to EmbSMOTE, with the performance gap widening as class imbalance increased. The study suggests that class-conditional structural fidelity, rather than surface-level diversity, is the key factor for effective augmentation, recommending retrieval-based oversampling as a default method. AI

IMPACT Suggests that current LLM-based text augmentation may not be suitable for imbalanced datasets, recommending classical methods instead.

RANK_REASON Academic paper presenting a benchmark and findings on text augmentation methods. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• EmbSMOTE
• GoEmotions-28
• Qwen3_8B

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →