[Paper Review] Ontology-Grounded Topic Modeling for Climate Science Research

Jennifer Sleeman, Tim Finin+1 more|arXiv (Cornell University)|Jul 28, 2018

Advanced Text Analysis Techniques12 references3 citations

TL;DR

This paper proposes ontology-grounded topic modeling to improve interpretability and relevance in climate science research by grounding topic models in domain-specific concepts from an IPCC-derived ontology. By replacing standard bag-of-words modeling with ontology-defined phrases, the method generates more contextually accurate, semantically meaningful topics that better reflect scientific terminology and improve knowledge discovery, with lower perplexity and enhanced explainability in climate science text analysis.

ABSTRACT

In scientific disciplines where research findings have a strong impact on society, reducing the amount of time it takes to understand, synthesize and exploit the research is invaluable. Topic modeling is an effective technique for summarizing a collection of documents to find the main themes among them and to classify other documents that have a similar mixture of co-occurring words. We show how grounding a topic model with an ontology, extracted from a glossary of important domain phrases, improves the topics generated and makes them easier to understand. We apply and evaluate this method to the climate science domain. The result improves the topics generated and supports faster research understanding, discovery of social networks among researchers, and automatic ontology generation.

Motivation & Objective

• To address the challenge of low interpretability and relevance in standard topic modeling for scientific texts, particularly in climate science.
• To improve topic modeling performance by grounding it in domain-specific ontologies derived from authoritative sources like IPCC glossaries.
• To support faster research understanding, discovery of researcher networks, and automatic ontology generation in climate science.
• To demonstrate that ontology-grounded topic modeling yields more fine-grained, context-specific topics than standard LDA or n-gram approaches.
• To enable better integration with knowledge bases and question-answering systems for scientific domains.

Proposed method

• Extract key phrases from IPCC report glossaries to seed a domain-specific ontology.
• Use the ontology to guide phrase extraction, replacing generic n-grams with semantically meaningful, domain-specific phrases.
• Apply Latent Dirichlet Allocation (LDA) with the ontology-grounded vocabulary instead of a standard bag-of-words model.
• Train the topic model on two domains: IPCC report chapters and cited scientific research papers.
• Use perplexity as a quantitative metric to evaluate model performance against standard LDA and n-gram baselines.
• Link generated topics to ontology concepts to support knowledge base population and explainability.

Experimental results

Research questions

• RQ1 Can grounding topic models with domain-specific ontologies improve topic interpretability and relevance in climate science texts?
• RQ2 How does ontology-grounded topic modeling compare to standard LDA and n-gram-based topic modeling in terms of topic quality and perplexity?
• RQ3 To what extent do ontology-grounded topics better reflect scientific terminology and context compared to standard approaches?
• RQ4 Can this method support automatic ontology generation and enhance knowledge base construction in scientific domains?
• RQ5 How does the use of domain-specific phrases from glossaries improve topic modeling in highly technical scientific literature?

Key findings

• Ontology-grounded topic modeling produced topics with significantly improved interpretability and context specificity compared to standard LDA and n-gram baselines.
• The method achieved lower perplexity on the IPCC dataset, indicating better fit to the data and improved topic coherence.
• Topics generated using ontology-derived phrases such as 'radiative forcing' and 'sea level rise' were more semantically meaningful and directly aligned with scientific concepts.
• The approach enabled better alignment with external knowledge sources, as evidenced by higher relevance of search results when querying topic phrases.
• The integration of ontology concepts enhanced the potential for downstream applications such as knowledge base population and question-answering systems in climate science.
• The use of domain-specific glossaries as a source for phrase grounding proved effective and scalable for constructing reliable topic model vocabularies.