OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Models

Hierarchytransformers

HierarchyTransformers is a Python library that uses transformer encoder language models to represent hierarchies as embeddings in hyperbolic space. It provides the HiT model, a customized trainer, and documented methods for encoding entity names and probing subsumption relationships.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Hierarchytransformers — OpenViST project profile in AI Models.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use it when a Python workflow needs language-model-based embeddings for hierarchical entities or needs to explore subsumption scoring in hyperbolic space.

Who it is for

  • Developers building hierarchy-aware representation workflows with language models.
  • Researchers working with taxonomies, entity hierarchies or subsumption prediction.
  • Sentence-transformers users who want to explore hierarchy encoders or train HiT models.

Documented uses

  • Encode entity names as hierarchy-aware embeddings.
  • Compare child and parent embeddings to calculate subsumption scores.
  • Train or evaluate hierarchy encoder models using the project’s scripts and customized HiT trainer.
  • Extract hierarchies from source data and construct datasets with the documented DeepOnto dependency.
Before you choose

Requirements & limits

The repository is public and states an Apache-2.0 license. The README documents installation through PyPI or GitHub and requires Python 3.9 or later.

  • HiT represents hierarchy structure in a Poincaré ball using hyperbolic geometry.
  • The main model extends the sentence-transformers architecture, and the project describes alignment with sentence-transformers.
  • The documented dependencies include DeepOnto for hierarchy extraction and dataset construction, and Geoopt for hyperbolic arithmetic.
  • The README flags an evaluation issue with sentence-transformers 3.3.1 and describes 3.4.0.dev0 as the fix in its dependency guidance.
  • Subsumption scoring uses a validation-set-derived weighting and threshold rather than fixed general-purpose values.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Probe subsumption between encoded entities

Calculate subsumption scores for paired child and parent entity embeddings.

Before you start

  • A loaded HierarchyTransformer model
  • Child and parent entity names to compare
  • A validation-set-derived centri_score_weight and overall threshold

Steps

  1. Encode the child and parent entity names separately with convert_to_tensor=True.

  2. Calculate hyperbolic distances between the paired embeddings using model.manifold.dist.

  3. Calculate each embedding’s distance from the origin using model.manifold.dist0.

  4. Apply the documented empirical score: -(distance + centri_score_weight * (parent norm - child norm)).

Expected outcome

Subsumption scores for the entity pairs, using the empirical function described in the project documentation.

Guide limitations

  • The weighting and overall threshold are determined on a validation set; the README does not supply universal values.
  • The example demonstrates scoring and does not establish accuracy for a particular hierarchy.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
15 October 2026
Original website title
KRR-Oxford/HierarchyTransformers · Official homepage title · retained in its source language
How we review sourcesSuggest a correction