Sentiment analysis systems often stumble when the data they learn from is lopsided, and a new study tackles that weakness head-on. In research published in Knowledge and Information Systems, Hossein Nekouei and Mohammad Hossein Moattar introduce a framework that combines two rounds of data augmentation with a dual-expert model architecture to classify the sentiment of short social media posts drawn from heavily imbalanced datasets. The work targets a familiar problem in machine learning: when one class dominates a dataset, models tend to overfit to the majority category and perform poorly on the rarer ones.
The researchers focused on the Twitter US-Airline Sentiment dataset, which is skewed toward negative opinions at roughly 63 percent, with neutral posts making up about 21 percent and positive posts only about 16 percent. To give underrepresented classes more training examples, the team applied a first stage of augmentation built on paraphrasing and back-translation, using Italian as an intermediary language. In the first stage of augmentation, the team applied paraphrasing and back-translation, using Italian as an intermediary language.
In the second stage of augmentation, the original texts were modified by wrapping key emotional aspects in special tokens, marked as [start aspect] and [end aspect]. This tagging step is designed to help the model better comprehend which parts of a short, nuanced post carry the emotional signal. The approach addresses a challenge the authors describe as often overlooked: social media texts are short, laden with sarcasm, emojis, and context-dependent meanings, and traditional methods frequently fail on them due to overfitting and high computational demands.
Once the augmented data are prepared, they flow through a pipeline with two specialized components. A T5 text-to-text generative model serves as Expert 1, performing unsupervised extraction of key emotional aspects from the text. A RoBERTa-base model then acts as Expert 2, handling sentiment extraction. A SoftMax layer completes the pipeline by producing the final classification.
The models were fine-tuned under a rigorous experimental setup employing cross-validation with both fivefold and tenfold configurations. Evaluation covered three datasets: Yelp, Twitter US-Airline Sentiment, and the StanfordNLP IMDB Review dataset. The headline result comes from the Twitter dataset, where, using tenfold cross-validation over two epochs, the framework achieved 98.5 percent accuracy, 98.5 percent precision, 98.5 percent recall, and an F1-score of 98.5 percent. The baseline model, by contrast, plateaued at 92 percent for both accuracy and F1-score on the same data.
On the Twitter dataset, the baseline model plateaued at 92 percent for both accuracy and F1-score, while the proposed framework reported 98.5 percent for those metrics. The authors attribute the improvement to the combination of two-stage data augmentation, specialized tokenization of emotional aspects, and the dual-expert design. The models were fine-tuned under a rigorous experimental setup employing cross-validation with both fivefold and tenfold configurations, and the headline Twitter results were reported using k=10 over 2 epochs.
The study’s scope is worth noting. Experiments covered three datasets: Yelp, Twitter US-Airline Sentiment, and the StanfordNLP/IMDB Review dataset, with the abstract reporting the detailed headline metrics for the Twitter data. The back-translation stage specifically routes text through Italian as an intermediary language.
The framework pairs two-stage data augmentation with a dual-expert architecture in which a T5 text-to-text generative model serves 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.
The authors state that the results highlight the effectiveness of two-stage data augmentation, specialized tokenization, and a dual-expert approach in robust sentiment analysis for imbalanced datasets. The authors describe the approach as tackling 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 paper was published as volume 68, article number 304 in Knowledge and Information Systems on 03 November 2026. The abstract reports the framework's results on the Twitter US-Airline Sentiment dataset, where it significantly outperformed the baseline model.
Subject of Research: A dual-expert deep learning framework with two-stage data augmentation for sentiment classification on imbalanced social media text datasets.
Article Title: Two-stage augmentation and dual-expert modeling for robust sentiment analysis on imbalanced texts
Article References: Nekouei, H., & Moattar, M. H. (2026). Two-stage augmentation and dual-expert modeling for robust sentiment analysis on imbalanced texts. Knowledge and Information Systems, 68(1), Article 304. https://doi.org/10.1007/s10115-026-02927-7
Image Credits: AI Generated
DOI: 10.1007/s10115-026-02927-7
Keywords: sentiment analysis, data augmentation, back-translation, imbalanced datasets, T5 model, RoBERTa, Twitter US-Airline Sentiment, cross-validation, natural language processing, machine learning, text classification, dual-expert architecture
News Source: Denise Maddox. (November 5, 2026). Two-Stage Text Augmentation and Dual-Expert AI Model Reaches 98.5% Accuracy on Imbalanced Twitter Sentiment Data. Scienmag.
