Hierarchy-Transformers/HiT-MiniLM-L6-WordNetNoun

A Hierarchy Transformer Encoder (HiT) model that explicitly encodes entities according to their hierarchical relationships.

Model Description

HiT-MiniLM-L6-WordNet is a HiT model trained on WordNet's noun hierarchy with random negative sampling.

Developed by: Yuan He, Zhangdie Yuan, Jiaoyan Chen, and Ian Horrocks
Model type: Hierarchy Transformer Encoder (HiT)
License: Apache license 2.0
Hierarchy: WordNet (Noun)
Training Dataset: Download wordnet.zip from Datasets for HiTs on Zenodo
Pre-trained model: sentence-transformers/all-MiniLM-L6-v2
Training Objectives: Jointly optimised on hyperbolic clustering and hyperbolic centripetal losses

Model Sources

Repository: https://github.com/KRR-Oxford/HierarchyTransformers
Paper: Language Models as Hierarchy Encoders

Usage

HiT models are used to encode entities (presented as texts) and predict their hierarhical relationships in hyperbolic space.

Get Started

Install hierarchy_transformers (check our repository) through pip or GitHub.

Use the code below to get started with the model.

from hierarchy_transformers import HierarchyTransformer
from hierarchy_transformers.utils import get_torch_device

# set up the device (use cpu if no gpu found)
gpu_id = 0
device = get_torch_device(gpu_id)

# load the model
model = HierarchyTransformer.load_pretrained('Hierarchy-Transformers/HiT-MiniLM-L6-WordNetNoun', device)

# entity names to be encoded.
entity_names = ["computer", "personal computer", "fruit", "berry"]

# get the entity embeddings
entity_embeddings = model.encode(entity_names)

Default Probing for Subsumption Prediction

Use the entity embeddings to predict the subsumption relationships between them.

# suppose we want to compare "personal computer" and "computer", "berry" and "fruit"
child_entity_embeddings = model.encode(["personal computer", "berry"], convert_to_tensor=True)
parent_entity_embeddings = model.encode(["computer", "fruit"], convert_to_tensor=True)

# compute the hyperbolic distances and norms of entity embeddings
dists = model.manifold.dist(child_entity_embeddings, parent_entity_embeddings)
child_norms = model.manifold.dist0(child_entity_embeddings)
parent_norms = model.manifold.dist0(parent_entity_embeddings)

# use the empirical function for subsumption prediction proposed in the paper
# `centri_score_weight` and the overall threshold are determined on the validation set
subsumption_scores = - (dists + centri_score_weight * (parent_norms - child_norms))

Training and evaluation scripts are available at GitHub. Technical details are presented in the paper.

Full Model Architecture

HierarchyTransformer(
  (0): Transformer({'max_seq_length': 128, 'do_lower_case': False}) with Transformer model: BertModel 
  (1): Pooling({'word_embedding_dimension': 384, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': True, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': False})
)

Citation

Preprint on arxiv: https://arxiv.org/abs/2401.11374.

Yuan He, Zhangdie Yuan, Jiaoyan Chen, Ian Horrocks. Language Models as Hierarchy Encoders. arXiv preprint arXiv:2401.11374 (2024).

@article{he2024language,
  title={Language Models as Hierarchy Encoders},
  author={He, Yuan and Yuan, Zhangdie and Chen, Jiaoyan and Horrocks, Ian},
  journal={arXiv preprint arXiv:2401.11374},
  year={2024}
}

Model Card Contact

For any queries or feedback, please contact Yuan He (yuan.he@cs.ox.ac.uk).