Knowledge and Information Systems Human-Computer Interaction

Two-stage augmentation and dual-expert modeling for robust sentiment analysis on imbalanced texts

Nekouei, Hossein, Moattar, Mohammad Hossein
Springer Nature · 2026-11-03 · DOI 10.1007/s10115-026-02927-7

Abstract

This paper introduces a novel framework for sentiment analysis on imbalanced datasets, leveraging a dual-expert architecture and two-stage data augmentation for enhanced performance. We first address the challenge of limited minority class representation in the highly imbalanced Twitter US-Airline Sentiment dataset (approximately 63% negative, 21% neutral, and 16% positive) by applying data augmentation through paraphrasing and back-translation using Italian as an intermediary language. In the second stage, we augment the original text by enveloping key emotional aspects with special tokens [start aspect] and [end aspect] to improve model comprehension. Our approach tackles the overlooked challenge of short, nuanced social media texts—laden with sarcasm, emojis, and context-dependent meanings—where traditional methods often fail due to overfitting and high computational demands. The augmented data are processed through a pipeline utilizing a T5 text-to-text generative model as Expert 1 for unsupervised key emotional aspect extraction, followed by a RoBERTa-base model as Expert 2 for sentiment extraction, and a SoftMax layer for classification. We fine-tune the models under a rigorous experimental setup, employing cross-validation with k=fivefold and k=tenfold. Experiments on three datasets—Yelp, Twitter US-Airline Sentiment, and StanfordNLP/IMDB Review—demonstrate the framework’s efficacy. Using k=10 over 2 epochs, our model achieves 98.5% accuracy, 98.5% precision, 98.5% recall, and an F1-score of 98.5% on the Twitter dataset, significantly outperforming the baseline model which plateaued at 92% for both accuracy and F1-score. This highlights the effectiveness of two-stage data augmentation, specialized tokenization, and a dual-expert approach in robust sentiment analysis for imbalanced datasets.

Read at publisher →