SentiBERT and Enhanced Bi-GRU for Weather-related Text Classification Using Lexical Features

Abstract

The growing volume of weather related content on social media platforms, especially Twitter, has highlighted the need for robust classification models that can handle noisy, ambiguous, and emotionally subtle language. However, existing models machine learning such as Support Vector Machines (SVM) often fail to effectively capture implicit sentiment and sequential context in short, real time texts. This study addresses the challenge of weather related text classification by proposing a hybrid architecture that combines SentiBERT, a sentiment aware transformer model, with an Enhanced BiGRU network equipped with Self Attention and LeakyReLU activation. Experiments were conducted using a five class(sunny, cloudy, rainy, extreme, other) dataset of weather related tweets with stratified cross validation across multiple deep learning models and tokenizers. Results show that the proposed SentiBERT + Enhanced BiGRU model outperformed all baselines, achieving 88.03% accuracy and 88.25% macro F1 score demonstrating its ability to better interpret contextual and emotional nuances. These findings imply that integrating sentiment specific embeddings with sequential modeling and lexical features offers a promising direction for future real time applications in climate monitoring and disaster alert systems.