Decoding Bias in #Climatechange Tweets: A Neural Language Style Transfer Approach

In response to the increase of biased and misleading opinions on social media regarding climate change, we present the Neural Language Style Transfer Bias (NLST Bias) framework—an AI-driven solution to identify and mitigate bias in climate-related tweets. Our approach analyses a curated dataset of climate change tweets collected from Twitter, leveraging spaCy for text pre-processing and BERT-based sentence embeddings to extract semantic features. Evaluation metrics with classifier models showed significant improvements, with CNN-LSTM achieving the highest F1 score of 78%. Most classifiers show marked improvements across accuracy, precision, recall, and F1 score with NLST, highlighting its effectiveness on this dataset. NLST Bias employs DistilBERT to classify tweets as biased or unbiased, with the Named Entity Recognition (NER) model trained on RobERTa transformers to identify word-level bias indicators. The framework's core component is a transformer-based style transfer model (e.g., GPT-2) that modifies the stylistic tone of biased text while preserving its semantic content. This transfer uses a style-specific decoder within the transformer architecture, ensuring the generated text maintains grammatical accuracy and contextual relevance. The resulting output exhibits tone, phrasing, and word choice adjustments to foster a more neutral style. Results indicate a significant reduction in bias without compromising text fluency or naturalness, verified through statistical metrics and human evaluations. Furthermore, Explainable AI techniques are integrated to enhance transparency in model decisions. Our findings underscore NLST Bias's efficacy in countering climate misinformation and promoting balanced discussions on social media, highlighting its potential as a tool for responsible AI in digital discourse.