Automated Construction and Expansion of Hierarchical Taxonomies

This direction is merged into another direction: manual taxonomy scalability. Its last version stays readable and citable here.

Version 1 of 151 papers51 conceptsfrozen 2026-10-04

In the modern information landscape, we rely on "knowledge backbones"—hierarchical taxonomies that organize everything from scientific disciplines and product categories to job titles and web search results. However, as noted by Man Abrol and Mani Abrol in their work on enterprise portals, manual curation cannot scale with the "firehose" of new data. When new concepts like "Scrum Master" or "Transformer Models" emerge, they must be integrated into existing structures immediately, or those structures become obsolete.

The research direction of Automated Taxonomy Construction and Expansion seeks to turn this manual bottleneck into an algorithmic process. By treating a taxonomy not just as a static list, but as a geometric structure or a linguistic map, researchers have developed ways to automatically grow, refine, and even generate these hierarchies from scratch.

The Geometry of Hierarchy

Why is it so hard to automate the placement of a new term in a tree? Traditional machine learning often represents words in Euclidean space (flat, multi-dimensional grids). However, trees are inherently non-Euclidean: the number of nodes grows exponentially as you move from the root to the leaves.

To solve this, Mingyu Derek Ma and colleagues introduced HyperExpan, which uses hyperbolic representation learning. In hyperbolic space, the "area" available expands exponentially with the radius, mirroring the structure of a tree. This allows models to embed thousands of concepts with minimal distortion. Similarly, Jie Huang proposed the concept of Semantic Capacity, using hyperbolic models to measure the "meaning scope" of a term—mathematically determining why "Artificial Intelligence" is a parent of "Linear Regression" based on their relative granularity.

Taking this further, Sahil Mishra and Avi Patni developed QuanTaxo, which represents entities as density matrices in a complex Hilbert space. This quantum-inspired approach handles "hierarchical polysemy"—the fact that a word's meaning shifts depending on which branch of the tree it occupies—by using superposition and interference to model overlapping relationships.

Learning from the Tree Itself

If you have a partial taxonomy, you already have a powerful training set. Modern frameworks use self-supervision to teach models how to "re-attach" existing nodes, effectively practicing for the arrival of new ones.

  • Path Ranking: Zichen Liu and Hongyuan Xu in TEMP treat expansion as a ranking problem. Instead of just looking at a single parent, the model evaluates entire "taxonomy paths" to find the most logical address for a new concept.
  • Mini-Paths: The STEAM framework, developed by Yue Yu and others, breaks the taxonomy into "mini-paths" (small parent-child-sibling clusters). By learning the local patterns of these clusters, the model can accurately predict where a novel term fits without needing to understand the entire global structure at once.
  • Multi-View Fusion: TaxoEnrich, by Minhao Jiang and Jiawei Han, looks at both "vertical" (ancestors) and "horizontal" (siblings) relationships simultaneously, ensuring that a new term is semantically similar to its brothers while being a logical subset of its father.

LLMs as Taxonomic Engines

The rise of Large Language Models (LLMs) has shifted the focus from training specialized classifiers to extracting the latent taxonomic knowledge already stored in these models.

Kunihiro Takeoka and his team developed Musubu, which treats BERT as an implicit knowledge base. By prompting the model with "Hearst patterns"—linguistic templates like "Fruit such as [Orange]"—they can extract hierarchical relationships even when domain-specific training data is scarce. Similarly, TaxoPrompt uses prompt tuning and random walks to capture the global context of a taxonomy, allowing the model to "generate" the correct position for a term rather than just classifying it.

In scientific domains, where terminology is highly specialized, Kun Zhu and Lizi Liao have shown that LLMs can guide multi-aspect clustering. By analyzing papers through different lenses—methodology, datasets, and objectives—the system creates more nuanced "family trees" than simple keyword matching ever could.

The Human in the Loop

Despite these advances, fully automated systems can sometimes lack the precision required by domain experts. This has led to a "middle ground" of human-in-the-loop systems.

Mihai Surdeanu and colleagues presented Taxonomy Builder, a tool that streamlines the construction process by allowing humans to guide the automated expansion. This mirrors the approach of Byron C Wallace, who used community-sourced constraints to build conference keyword taxonomies. These systems recognize that while machines are better at processing the "firehose" of data, humans are still the ultimate judges of semantic utility.

By combining geometric representations, self-supervised path learning, and the latent knowledge of LLMs, researchers are moving toward a world where our maps of knowledge grow as fast as the knowledge itself.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_automated_construction_and_expansion_of__cfc7f60e,
  title = {Automated Construction and Expansion of Hierarchical Taxonomies},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/automated-construction-and-expansion-of-hierarchical-taxonomies/rdr_cfc7f60e800147f1bec72b98eae090d4},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_automated_construction_and_expansion_of__cfc7f60e_v1,
  title = {Automated Construction and Expansion of Hierarchical Taxonomies},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/automated-construction-and-expansion-of-hierarchical-taxonomies/rdr_cfc7f60e800147f1bec72b98eae090d4/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}