Dataset for DAFN: A Dual Attention Fusion Network for Textual Data Classification

Abstract

Sentiment classification is challenging because short, noisy texts and longer reviews exhibit different linguistic characteristics. Tweets contain limited context, informal expressions, and abrupt polarity shifts, whereas reviews require effective modelling of long-range dependencies and stable optimisation over extended sequences. This paper proposes DAFN, a Dual Attention Fusion Network for binary sentiment classification across tweet and review datasets. DAFN combines dual representations through early fusion, processes them using parallel BiGRU and BiLSTM branches, and introduces an Attention Interaction module for cross-branch communication. It further integrates multi-scale temporal convolutions, hierarchical pooling, feature recalibration, and residual multi-level fusion to improve information flow and robustness across variable text lengths. Experiments on three tweet and five review datasets show that DAFN consistently outperforms earlier hybrid deep-learning baselines and remains competitive with strong transformer models. It achieves the best overall performance on Airline, T4SA, and App, strong AUC results across several review datasets, and near-parity with RoBERTa on Kindle. Computational analysis shows that the static variant is more efficient, while the contextual variant delivers stronger predictive performance. Overall, DAFN provides a unified and effective framework for sentiment classification across heterogeneous text lengths.