Tracking the evolution of scientific language over centuries presents a significant challenge for computational linguistics. This research introduces a novel approach using non-Euclidean geometry to map the "vertical" growth of scientific knowledge—from general concepts to highly specialized terminology—without relying on labor-intensive, hand-crafted ontologies.
The Challenge of Diachronic Ontologies
Language is inherently dynamic, yet most tools for mapping its structure are static. Resources like WordNet provide excellent hierarchical maps (e.g., identifying that a "proton" is a type of "subatomic particle"), but these are manually constructed for contemporary language. Creating similar maps for historical data—known as diachronic ontologies—is difficult because word meanings shift, new terms emerge, and old ones disappear. Standard word embeddings (like Word2Vec) excel at finding "horizontal" similarity (synonyms) but struggle to capture the "vertical" hierarchy of specialization.
Methodology: Hyperbolic (Poincaré) Embeddings
To address this, the researchers utilize Hyperbolic (Poincaré) embeddings. Unlike standard Euclidean geometry, hyperbolic space expands exponentially as one moves away from the center, making it mathematically ideal for representing tree-like hierarchies.
In this model, the distance from the origin (the "norm") serves as a proxy for specificity: general terms remain near the center, while specialized terms are pushed toward the periphery. To extract relationships, the authors used an "agnostic" method, treating words as related if they appeared in the same sentence. They divided the corpus into five 50-year periods, training a separate Poincaré model for each to observe temporal changes in semantic space.
Data and Domain Focus
The study utilized the Royal Society Corpus (RSC), a massive dataset containing approximately 10,000 journal articles from the Transactions and Proceedings of the Royal Society in London (1665–1869), totaling roughly 32 million tokens. To reduce noise, the analysis focused on four scientific domains: Chemistry, Physiology, Botany, and Astronomy. Modern WordNet was used as a benchmark to validate whether the calculated "norms" correlated with the number of dictionary senses a word has today.
Key Insights: The Geometry of Specialization
The most striking finding is that scientific maturation has a distinct geometric signature. As fields evolved, the "norm" of their vocabulary increased, indicating a literal expansion of semantic space as researchers moved from general observations to specialized nomenclature.
Key results include:
- Quantifying Hierarchy: In Astronomy, the average norm of words grew from 3.2 to 20.9 over the study period, reflecting a massive expansion of specialized terminology.
- WordNet Correlation: A significant negative correlation (up to -0.46) was found between a word's norm in the historical model and its number of senses in modern WordNet. This confirms the model successfully identifies specialized terms, which typically have fewer, more precise meanings.
- Uneven Evolution: Chemistry and Physiology showed stronger hierarchical organization over time compared to Botany or Astronomy, reflecting different speeds of systematic conceptualization across fields.
Limitations and Future Directions
While the agnostic co-occurrence approach works at scale, it can be noisy, occasionally linking words that lack a true hierarchical relationship. Furthermore, using modern WordNet as a benchmark does not account for words that have completely changed meaning since the 18th century. Future work may incorporate directional links to more explicitly define "parent-child" relationships in these automatically generated ontologies.
Summary updated