Abstract
The 2nd Symposium on NLP for Social Good (NSG 2024) convened an international cohort of researchers to explore the application of Natural Language Processing (NLP) to pressing global challenges. This report synthesizes the key contributions from the symposium, ranging from climate change discourse analysis and mental health support via Large Language Models (LLMs) to the automation of legal reasoning and the development of healthcare "Digital Twins." The findings underscore the potential of AI to advance the United Nations’ Sustainable Development Goals through interdisciplinary collaboration.
Introduction
Despite the rapid advancement of NLP technologies, a significant "application gap" persists between theoretical benchmarks and real-world societal impact. Much of the current innovation remains confined to commercial interests or abstract research. The NSG 2024 symposium aimed to bridge this gap by connecting NLP experts with practitioners in social sciences, law, and medicine. The goal is to ensure that technological tools are designed to solve humanitarian and environmental challenges effectively, rather than existing in a vacuum.
Methodological Approaches
The symposium featured several distinct methodologies for applying NLP to social good:
- Discourse Analysis & Framing: Research presented by Prof. Manfred Stede utilized clause-level annotation and automated classification to identify "framing"—the specific perspective or emphasis taken on an issue—within climate change editorials and political communications.
- Interpretable LLMs for Mental Health: Prof. Sophia Ananiadou introduced MentaLLaMA, a specialized series of LLMs designed for social media mental health analysis. This approach prioritizes interpretability, moving beyond binary sentiment analysis to evaluate sentiment intensity and emotional cues.
- Bridging Legal Theory and AI: Prof. Kevin D. Ashley explored the use of generative AI and text analytics to transform raw legal case texts into structured "argument models," making complex legal reasoning more accessible to professionals.
- Knowledge Representation for Digital Twins: Dr. Swarnendu Ghosh discussed methods for structuring heterogeneous healthcare data—including Electronic Health Records (EHRs), genomics, and clinical reports—to create "Digital Twins." These virtual patient representations facilitate predictive analytics and remote monitoring.
Data Sources
The research presented utilized a diverse array of datasets:
- Climate Discourse: A corpus comprising editorials from Nature and Science, German parliamentary speeches, press releases, and tweets spanning 2017–2021.
- Mental Health & Misinformation: Social media feeds were used for mental health analysis, alongside a novel conspiracy dataset developed for detecting misinformation.
- Legal & Healthcare: Data included legal case archives and heterogeneous healthcare sources such as EHRs and clinical guidelines.
Key Insights and Results
The symposium, which attracted 177 registered participants globally, yielded several significant findings:
- The Importance of Framing: In climate change communication, the framing of a message is as critical as its content. NLP tools can now quantify these frames across different channels, revealing how discourse varies between formal parliamentary records and social media platforms like Twitter.
- Affective Cues in Mental Health: Simple sentiment analysis is insufficient for mental health applications. Incorporating affective (emotional) cues significantly improves the accuracy of automated systems in detecting misinformation and understanding psychological states.
- Generative AI in Law: While representing legal knowledge was traditionally a manual process, generative AI is now being used to automate the extraction of legal "factors" and "rules." However, integrating these symbolic rules with sub-symbolic deep learning models remains a primary challenge.
- Healthcare Interoperability: The development of "Digital Twins" is currently hindered not by a lack of data, but by "isolated innovation"—the inconsistency of data structures across different platforms.
Challenges and Future Directions
The report identifies several hurdles for the field:
- Symbolic vs. Sub-symbolic Integration: There is a persistent technical challenge in merging domain-specific symbolic knowledge (e.g., legal rules) with deep learning models.
- Data Standardization: Large-scale deployment of healthcare AI requires standardized data structures to overcome current fragmentation.
- The Evolving Role of Models: As generative AI advances, researchers must determine whether AI will replace traditional argument models or serve as a tool to build them.
- Empowerment through Collaboration: Future work must focus on ensuring that NLP tools empower the social sector rather than merely observing it, requiring sustained interdisciplinary partnership.
Summary updated