From LSA to LLM: Evolution and limitations of topic modelling methods for biodiversity conservation

Biodiversity conservation faces mounting challenges due to accelerating species loss driven by habitat degradation, agricultural expansion, and climate change. Scientific literature grows rapidly, making automated text analysis tools increasingly essential for synthesizing knowledge and informing policy. This study reviews the evolution of topic modelling techniques—from Latent Semantic Analysis (LSA) and Latent Dirichlet Allocation (LDA) to embedding-based models like BERTopic and Top2Vec, and most recently, Large Language Models (LLMs) such as GPT. While traditional probabilistic models remain valuable for identifying thematic structures in ecological texts, LLMs offer enhanced flexibility, particularly for grey literature and prompt-based analysis. Applications of topic modelling in biodiversity conservation include identifying knowledge gaps, monitoring public perceptions, analysing policy texts, and summarising scientific literature. Despite recent advances, key challenges persist, including lack of domain-specific models, limited access to ecological corpora, and reproducibility concerns with AI models. Furthermore, the environmental cost of large-scale computation must be weighed against ecological goals. This review highlights that while AI-based topic models present valuable tools for ecological research and evidence synthesis, their role should remain complementary to human expertise and field-based knowledge. • Techniques have progressed from classical linear algebra methods like LSA to probabilistic models (LDA, STM) and embedding-based or generative models such as BERTopic, Top2Vec, and TopicGPT. • Topic modelling supports literature synthesis, stakeholder engagement, policy analysis, monitoring conservation projects, and identifying emerging threats or knowledge gaps. • Despite increasing use, topic modelling is still relatively rare in biodiversity research, especially before 2010, indicating untapped potential for guiding evidence-based conservation strategies. • While LLMs enable more flexible topic discovery, they have limitations, require human oversight, have high energy costs, and lack transparency compared to traditional probabilistic models. • AI-based topic models enhance research efficiency and insight extraction but cannot replace field knowledge, address systemic biodiversity data gaps, or substitute for evidence-based conservation action.