OntoZSL

Code and Data for the paper: "OntoZSL: Ontology-enhanced Zero-shot Learning". Yuxia Geng, Jiaoyan Chen, Zhuo Chen, Jeff Z. Pan, Zhiquan Ye, Huajun Chen and others. The Web Conference (WWW) 2021 Research Track.

In this work, we propose to utilize ontology and generative adversarial network to deal with the zero-shot learning problems in image classification and KG completion.

Ontological Schemas

Dataset Description

IMGC

Dataset	# Classes (Total/Seen/Unseen)	# Ontology Schema (Triples/Concepts/Properties)
AwA	50 / 40 / 10	1,256 / 180 / 12
ImNet-A	80 / 28 / 52	563 / 227 / 19
ImNet-O	35 / 10 / 25	222 / 115 / 8

KGC

Dataset	# Relations (Total/Train/Val/Test)	# Ontology Schema (Triples/Concepts/Properties)
NELL-ZS	139 / 10 / 32	3,055 / 1,186/4
Wikidata-ZS	469 / 20 / 48	10,399 / 3,491/8

Requirements

python 3.5
PyTorch >= 1.5.0

Dataset Preparation

Word Embeddings

You need to download pretrained Glove word embedding dictionary, uncompress it and put all files to the folder data/glove/.

AwA2

Download public image features and dataset split for AwA2, uncompress it and put the files in AWA2 folder to our folder data/AwA2/.

ImageNet (ImNet-A, ImNet-O)

Download the image features and the word embeddings of ImageNet classes as well as their splits from here and put them to the folder data/ImageNet/.

NELL-ZS & Wikidata-ZS

You can download these two datasets from here and put them to the corresponding data folder.

OntoZSL Training

The first thing you need to do is to train the text-aware ontology encoder using the code in the folder code/OntoEncoder, you can get more details at code/OntoEncoder/README.md.

Secondly, with well-trained ontology embedding, you can take it as the input of generative model, see the codes in the folders code/IMGC and code/KGC. The running commands are listed in the corresponding README.md files.

Note: you can skip the first step if you just want to use the ontology embedding we learned, the files are provided in the corresponding directories.

How to Cite

If you find this code useful, please consider citing the following paper.

@inproceedings{geng2021ontozsl,
  author    = {Yuxia Geng and
               Jiaoyan Chen and
               Zhuo Chen and
               Jeff Z. Pan and
               Zhiquan Ye and
               Zonggang Yuan and
               Yantao Jia and
               Huajun Chen},
  editor    = {Jure Leskovec and
               Marko Grobelnik and
               Marc Najork and
               Jie Tang and
               Leila Zia},
  title     = {OntoZSL: Ontology-enhanced Zero-shot Learning},
  booktitle = {{WWW} '21: The Web Conference 2021, Virtual Event / Ljubljana, Slovenia,
               April 19-23, 2021},
  pages     = {3325--3336},
  publisher = {{ACM} / {IW3C2}},
  year      = {2021},
  url       = {https://doi.org/10.1145/3442381.3450042},
  doi       = {10.1145/3442381.3450042},
  timestamp = {Thu, 14 Oct 2021 10:04:23 +0200},
  biburl    = {https://dblp.org/rec/conf/www/GengC0PYYJC21.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

Acknowledgement

We refer to the code of LisGAN and ZSGAN. Thanks for their contributions.

genggengcss/OntoZSL