Introduction
In the fields of Earth science and oceanography, scientific variables are rarely simple "atomic" entities like gene names or proteins. Instead, they are often complex, nested noun phrases that incorporate state, location, and timing—for example, "the timing and magnitude of surface temperature evolution in the Southern Hemisphere." This complexity leads to high sparsity in literature; specific descriptions rarely repeat exactly across different studies, making it difficult for standard text-mining tools to identify relationships between papers.
Most existing text mining tools were developed for the biomedical field, where entities are often proper nouns. In contrast, Earth science variables describe a state and its context. Because these descriptions are so specific, over 66% of extracted variables in a typical corpus are unique. Researchers from NTNU have developed a system to automatically extract these variables and generalize them into a searchable hierarchy, addressing the challenge of how to recognize when different long phrases refer to the same underlying concept.
Methodology
Variable Extraction
The system employs Tregex, a syntax-aware pattern-matching tool, to identify variables based on their grammatical structure rather than simple keywords. By defining patterns within syntactic trees, the system can capture not only the variable itself but also the direction of variation (increase, decrease, or change). This approach prioritizes scientific "events" over static entities, recognizing that in Earth science, the change is often more important than the thing itself.
Variable Generalisation
To address the sparsity problem, the authors use Tsurgeon to perform "progressive pruning" of syntax trees. This iterative process strips away modifiers—such as initial determiners, adjectives (premodifiers), and prepositional phrases (postmodifiers)—to transform specific mentions into broader "parent" concepts. For example, "annual temperature between 1958 and 2010" might be generalized to "annual temperature" and eventually to "temperature." This syntactic pruning allows the system to discover broader categories naturally without requiring a pre-existing ontology.
Dataset and Implementation
The system was tested on the Nature Marine Science Corpus, a curated dataset of approximately 10,000 abstracts from journals published by the Nature Publishing Group (post-1997) related to marine science and oceanography. The final processed corpus contains 9,586 abstracts, roughly 60,000 sentences, and 4 million tokens. The authors have made both the corpus and the extracted variables publicly available to support further research.
Key Results
The extraction process yielded 21,817 matched variables, categorized by direction: 9,352 for "change," 7,400 for "increase," and 5,065 for "decrease." The generalization step was particularly impactful; while the majority of extracted variables were initially unique, the pruning rules generated over 150,000 potential variations. By filtering for those appearing at least twice, the researchers established a hierarchy of 17,613 variable types.
The authors developed a browsable hierarchy that allows researchers to click on a general term (like "annual temperature") and see all the specific, complex mentions of it across thousands of papers. This interface uses color-coding to indicate whether the variable was increasing or decreasing in the original text, providing a powerful tool for meta-analysis and data discovery.
Discussion and Future Work
While effective, the system is sensitive to the accuracy of the initial syntactic parse. Ambiguities in "PP-attachment" (deciding which noun a prepositional phrase modifies) can lead to incoherent extractions. Furthermore, the current model does not account for semantic nuances where change is implicit (e.g., "ocean acidification" implies increasing acidity) or handle co-reference (e.g., pronouns like "it" referring back to a variable).
Future work aims to move toward Structural Modeling, linking variables to specific geographic locations and timeframes, and mining for causal relations between different variables (e.g., how an increase in greenhouse gas concentrations causes a decrease in sea-ice cover).
Summary updated