Explainable Multitask Learning for Sentiment and Emotion Classification in Reddit-Based Public Relations Crisis Management

Hon, Edwina Kai Xin (2026) Explainable Multitask Learning for Sentiment and Emotion Classification in Reddit-Based Public Relations Crisis Management. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text RDS_Edwina Hon Kai Xin_Fulltext.pdf Restricted to Registered users only Download (1MB)

Abstract

Social media crises such as the Samsung Galaxy Note 7 recall unfold rapidly and generate intense public reactions. To respond effectively, organisations require tools that not only detect sentiment but also explain the reasoning behind model predictions. Current sentiment analysis approaches often prioritise accuracy over interpretability and struggle to adapt to domain-specific crisis contexts. This project develops an explainable multitask learning (MTL) framework for sentiment and emotion classification in Reddit-based crisis communication. Transformer models (BERTweet and DistilRoBERTa) were fine-tuned in both single-task and multitask configurations, trained on benchmark datasets and evaluated on a curated Reddit corpus from the Note 7 crisis. Integrated Gradients (IG) was adopted as the explanation method, assessed against faithfulness, stability, and fairness criteria. Results show that single-task models consistently outperform MTL in general-domain benchmarks, while MTL offers modest improvements for sentiment classification in crisis-specific data. Emotion classification remains challenging, with consistently low performance across configurations. To ensure robustness, models were evaluated under both single-seed and five-seed protocols, with the latter providing more conservative and reproducible estimates of stability. IG explanations provided transparency and fairness but suffered from poor stability and limited faithfulness. Overall, the study highlights both the promise and limitations of explainable multitask learning for crisis informatics. It contributes a framework that balances predictive accuracy with interpretability, offering a foundation for more trustworthy AI tools to support transparent and effective crisis communication.