Science Technology Ontology: A Taxonomy of Emerging Topics

Mahender Kumar, Ruby Rani, Mirko Bottarelli, Gregory Epiphaniou, Carsten MapleTEXT2KG/BiKE@ESWC 2023 · OpenReview

Abstract

Traditional scientific ontologies are often siloed and slow to update, failing to capture the rapid emergence of interdisciplinary research. This paper presents the Science and Technology Ontology (S&TO), an automated system leveraging BERTopic and large-scale scholarly data from the Semantic Scholar Academic Graph. S&TO identifies over 5,000 distinct topics across computer science, physics, chemistry, and engineering, providing a dynamic, self-updating map of the modern research landscape.

Introduction

Ontologies serve as the foundational maps for navigating scientific knowledge. However, established systems like the ACM Computing Classification System or Medical Subject Headings (MeSH) are manually curated, making them inherently static and domain-specific. As research increasingly crosses traditional boundaries, these systems struggle to categorize "unconventional" or emerging fields. The Science and Technology Ontology (S&TO) addresses this gap by automating the discovery and categorization of research topics using machine learning, ensuring that the "map" of science evolves as quickly as the research itself.

Methodology

The S&TO framework utilizes a sophisticated pipeline based on BERTopic to process scientific abstracts into a structured taxonomy. This unsupervised approach allows for the discovery of topics without prior human labeling.

The five-stage machine learning pipeline used to generate the S&TO, from initial embedding to final topic representation.

The workflow consists of five primary stages:

  1. Sentence Embedding: Abstracts are converted into high-dimensional vectors using the paraphrase-MiniLM-L12-v2 model to capture semantic meaning.
  2. Dimension Reduction: UMAP (Uniform Manifold Approximation and Projection) is employed to project these vectors into a lower-dimensional space while preserving local and global structures.
  3. Clustering: HDBSCAN identifies dense clusters of related papers. Its density-based nature is crucial for filtering out noise and identifying clusters of varying shapes.
  4. Vectorization & Representation: c-TF-IDF (Class-based Term Frequency-Inverse Document Frequency) identifies the most descriptive keywords for each cluster.
  5. Graph Construction: The resulting topics and their relationships (e.g., hierarchy and similarity) are stored in a Neo4j graph database using the SKOS (Simple Knowledge Organization System) data model.

Data and Scale

The ontology was built using a dataset of 393,991 scientific articles retrieved via the Semantic Scholar Academic Graph (S2AG) API, spanning October 2021 to August 2022. The current iteration includes 5,153 distinct topics and 13,155 semantic relations, providing a high-resolution view of the research landscape.

Results and Taxonomy Structure

A key strength of S&TO is its ability to capture the hierarchical and interdisciplinary nature of emerging research. By carefully tuning clustering parameters, the model maintains a low outlier rate (15.47%), ensuring that even niche, "unconventional" topics are represented rather than discarded as noise.

A hierarchical representation of topics generated by the S&TO, demonstrating how the system organizes specific research areas into a tree-like taxonomy.

The system calculates the "strength of collaboration" between topics using a harmonic mean of paper probabilities. This allows researchers to see not just what the topics are, but how closely they interact. This provides a significantly higher resolution than standard keyword searches, mapping specific sub-topics like "motorized_spindle_aerostatic_restrictors" within broader engineering categories.

Discussion and Future Work

While S&TO provides a powerful tool for automated knowledge mapping, challenges remain. The unsupervised nature of BERTopic can lead to clunky, automatically generated labels that lack the clarity of human-written category names. Furthermore, the current model is limited to four major scientific fields. Future work will focus on expanding domain coverage to include biology and environmental sciences, improving label quality through manual tagging, and exploring advanced clustering algorithms to further refine the accuracy of the knowledge landscape.

Summary updated

BibTeX
@inproceedings{kumar2023science,
  author = {Mahender Kumar and Ruby Rani and Mirko Bottarelli and Gregory Epiphaniou and Carsten Maple},
  title = {{Science Technology Ontology: A Taxonomy of Emerging Topics}},
  booktitle = {TEXT2KG/BiKE@ESWC},
  year = {2023},
  url = {https://openreview.net/forum?id=bpag7w87aS},
}
Figure