ICON

ICON (Implicit CONcept Insertion) is a self-supervised taxonomy enrichment system. It discovers implicit intermediate concepts missing from an existing taxonomy by:

1. Retrieving a cluster of related concepts via an embedding model (emb_model)
2. Generating a virtual concept label for each subset via a generative model (gen_model)
3. Placing the concept via subsumption-prediction-based search (sub_model)

Dependencies

Core: numpy, rdflib, networkx, faiss, tqdm, nltk
Sub-model training: torch, pandas, transformers, datasets, evaluate, info-nce-pytorch
Python ≥ 3.9

Quick start

from icon import ICON, from_json

taxo = from_json("data/raw/google.json")
icon = ICON(data=taxo, emb_model=emb, gen_model=gen, sub_model=sub)
icon.run()  # auto mode — enriches entire taxonomy

See demo.ipynb for a full walkthrough including model wrapping.

Sub-models

ICON requires three plug-in models with these signatures:

Model	Signature	Purpose
emb_model	(query: List[str], ...) -> np.ndarray	Embed concepts for retrieval
gen_model	(labels: List[str], ...) -> str	Generate union label for a subset
sub_model	(sub, sup, ...) -> np.ndarray	Predict subsumption probability

emb_model and gen_model are optional depending on the operating mode (see configuration).

eBay models: Models fine-tuned on eBay data are available at RNO HDFS: /user/jingcshi/ICON_models/.

We provide a fine-tuning pipeline using BERT (emb/sub) and T5 (gen). See fine-tuning data reference and the notebooks under experiments/model_training/.

Documentation

• Configuration reference — all ICON parameters explained
• Fine-tuning data reference — data_config.json parameters for each sub-model
• Taxonomy file format — JSON schema for reading/writing taxonomies

Citation

@inproceedings{10.1145/3589334.3645584,
  author    = {Shi, Jingchuan and Dong, Hang and Chen, Jiaoyan and Wu, Zhe and Horrocks, Ian},
  title     = {Taxonomy Completion via Implicit Concept Insertion},
  year      = {2024},
  publisher = {Association for Computing Machinery},
  booktitle = {Proceedings of the ACM on Web Conference 2024},
  pages     = {2159--2169},
  series    = {WWW '24}
}