Class-Aware Semantic Hybrid Data Augmentation for Imbalanced Sentiment Classification Using Multiple Transformer Models
Year: 2026
Type: preprint
Abstract: begin{abstract}Class imbalance remains one of the major challenges in sentiment classification because transformer-based models are often biased toward majority classes, leading to poor recognition of minority sentiments. Although existing data augmentation techniques increase the number of training samples, they generally apply identical transformations across all classes without considering class-specific characteristics, which may introduce semantic drift, redundant samples, and limited improvements in minority-class learning.To address these limitations, this paper proposes a Class-Aware Semantic Hybrid Data Augmentation (CSHDA) framework for imbalanced sentiment classification. The proposed framework combines class-aware augmentation probabilities with four complementary augmentation techniques, namely BERT-based paraphrasing, synonym replacement, random insertion, and random deletion. To ensure the quality of generated samples, semantic consistency is validated using Sentence-BERT cosine similarity, while duplicate-aware filtering removes redundant augmented instances. The framework augments only minority classes, thereby preserving the original class distribution characteristics while reducing learning bias. The proposed framework was evaluated using BERT, RoBERTa, DistilBERT, and ALBERT transformer-based architectures on two benchmark datasets with different characteristics: the CUFC corpus and the Twitter Airline dataset. Experimental results demonstrate consistent improvements across both datasets. On the CUFC dataset, Macro F1-score improved from 60.80% to 81.10% for BERT and from 55.43% to 80.01% for DistilBERT. On the Twitter dataset, BERT improved from 60.65% to 69.82%, while DistilBERT increased from 42.98% to 71.21%. The proposed framework also substantially improved minority-class recall, particularly for the neutral sentiment class, while paired statistical analysis confirmed that the observed improvements were statistically significant. These results demonstrate that the proposed CSHDA framework effectively mitigates class imbalance by generating semantically consistent and diverse minority-class samples, leading to improved robustness and cross-dataset generalization of transformer-based sentiment classification models.\end{abstract}.
Source: Research Square
Authors: Prashant Upadhyaya, G.L. Saini, Atul Makrariya
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204465670