Information retrieval and multimodal sentiment analysis for climate change

• Main Author: Duong, Cuc Thi Kim
• Other Authors: Erik Cambria
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Opinion mining; Sentiment analysis; Climate change; Large language models; Twitter; Topic modeling; Social media
• Online Access: https://hdl.handle.net/10356/178373
• Institution: Nanyang Technological University

The Intergovernmental Panel on Climate Change (IPCC) has affirmed that human activities persist in elevating the global mean temperature, altering extreme weather patterns like heatwaves, droughts, tropical cyclones, and heavy rainfall. These climatic modifications pose significant economic and social threats if not promptly and effectively addressed. Governmental entities recognize the significant impacts of climate change on the environment and society, understanding that mitigating these effects necessitates urgent and robust action through public policy. However, the often time-consuming process of policy creation can potentially be expedited by leveraging social media to amass public opinions despite the challenges of unstructured data. This thesis critically examined public perspectives on climate change, particularly by leveraging Twitter data to highlight the value and practicality of social media in capturing real-world sentiment. Public reactions to the Australian wildfires, explored in Chapter 5, reinforced social media's capacity to reflect real-world events, aligning with previous research findings. This study further underscores the relevance of social media data in understanding public sentiment and encourages its continued use in similar contexts. In Chapter 3, the ClimateTweets dataset was created to evaluate various AI models for topic modeling and sentiment analysis. This dataset, a valuable resource for researchers, provides labeled data to assist in AI model development and evaluation. Using the dataset, Chapter 4 identified the optimal generative AI model for category and sentiment classification tasks. The results from Chapter 5 suggested that a hybrid approach using supervised and lexicon-based learning for sample extraction could surpass conventional methodologies in sentiment and emotion classification tasks, demonstrating the practical implications of this research. Investigating public perspectives on the climate change domain yielded important findings. The corpus of the ClimateTweets dataset provides initial insights about major topics and how polarized they are during the one-year time frame of data collection. In addition, it found that public opinions on social media contain harmful noises; therefore, users should be cautious and critical when reading these posts. When conducting a pilot study in mining public opinion about climate change-related hazards, such as wildfires in Australia, we found that people started to link wildfires as one of the impacts of climate change in the years 2016-2017. The chapter also found strong evidence that the trends of tweets can reflect the damage of wildfires in Australia, including the burned areas and the number of damaged buildings. Chapter 6 conducted another pilot study on mining public perspectives on important mitigation and adaptation strategies over six months of data, from July 2022 to December 2022. The results showed that the public mostly supported the strategies on energy, agriculture, forestry, and other land use topics. For most strategies, the number of opinions is proportional to their contribution to reducing global emissions. Importantly, public perception generally aligns with the IPCC's recommendations, providing reassurance about the accuracy and relevance of this study's findings.