Large language models in aquatic risk assessment: research status and future perspectives
Qianhui Li, Fei Cheng, Jing You · Environmental and Biogeochemical Processes · 2026
As a crucial component for maintaining ecological security and human health, aquatic ecosystems are facing risks from intensified human activities. Aquatic risk assessment requires a comprehensive understanding of geographic distribution, exposure, and effects of diverse pollutants. In the era of big data, utilizing available environmental data to its fullest extent is expected to facilitate efficient regional risk assessment, and support informed decision-making in risk management. However, it faces a significant challenge in data integration, as environmental data are scattered across heterogeneous texts from diverse corpora, such as scientific research literature, monitoring reports, and policy documents. Natural language processing (NLP) approaches serve as key tools for structured information extraction (IE). Traditional NLP techniques face bottlenecks such as cumbersome feature engineering, and limited generalization, while newly developed large language models (LLMs) can perform a wide array of tasks through prompting, achieving remarkable generalization and versatility. The present work systematically reviewed cutting-edge applications of LLMs in IE tasks across multiple disciplines, including chemistry, biology, and toxicology, from three perspectives: entity extraction, relation extraction, and semantic generation. On the contrary, the current application of LLMs in environmental science is still in its early stages, facing challenges such as data dependence, hallucinations, and environmental concerns. Future research should focus on building high-quality environmental corpora and hybrid strategies to systematically integrate aquatic ecological risk data, and support environmental risk assessment and management policies.