davendw49 / gakg

GAKG is a multimodal Geoscience Academic Knowledge Graph (GAKG) framework by fusing papers' illustrations, text, and bibliometric data.

GAKG: A Multimodal Geoscience Academic Knowledge Graph 😈

The SPARQL Query Endpoint of GAKG is gakg/sparql, with the Graph IRI: https://www.acekg.cn/.

Overview

GAKG is a multimodal Geoscience Academic Knowledge Graph (GAKG) framework by fusing papers' illustrations, text, and bibliometric data. To our knowledge, GAKG is currently the largest and most comprehensive geoscience academic knowledge graph, consisting more than 68 million triples. Figure 1 shows the overview of GAKG. And if you want to explore the entire GAKG, view https://gakg.acemap.info.

Figure 1. Overview of Multimodal GeoScience Academic Knowledge Graph (GAKG)

In order to better serve the data mining and knowledge discovery communities, GAKG preserves several datasets and GAKG-oriented resources.

No.	Resource Name	Resource Type	Link
1	GAKG Datasets	data dump	info, Google Drive, ftp
2	GA16K	Knowledge Representation Benchmark	info, /benchmarks/GA16K/, ftp
3	GPCN	Geoscience Papers Citation Network	info, ftp
4	GACN	Geoscience Authors Cooperation Network	info, ftp
5	GAKG SPARQL Endpoint	Query Endpoint	GAKG Snorql
6	Pipeline Supplements	models and codes	Google Drive

Community Detection

• Statistics

dataset	node	edge	num. lcc	clustering	triangle	daglongest
GPCN	842,121	16,034,510	1,450	0.0699	38,789,469	176
GPCN-lwcc	838,219	16,031,892	1	0.0701	38,789,345	176
GACN	860,280	5,381,861	32,609	0.690	43,502,542	15
GACN-lc	752,718	5,231,507	1	0.700	43,332,307	15

• Baselines

Baseline	Source Code	Paper
Louvain	https://sites.google.com/site/findcommunities/	Fast unfolding of communities in large networks 1
Map Equation	https://www.mapequation.org/code.html	Maps of random walks on complex networks reveal community structure 2
LPA	https://github.com/zhuo931077127/LPA-algorithm-Demo	Near linear time algorithm to detect community structures in large-scale networks 3

The entire benchmarks also can be accessed from google drive

Knowledge Representation Learning

• Statistics

Benchmark	Entities Type	Relations TypeS	Triples
GA16K	10	16,363	151,662
WN18	18	40,943	141,442
FB15K	1,345	14,951	483,142

• Baselines

Baseline	Source Code	Paper
RESCAL	https://github.com/mnick/rescal.py	A Three-Way Model for Collective Learning on Multi-Relational Data 4
TransE	https://github.com/zqhead/TransE	Translating Embeddings for Modeling Multi-relational Data 5
TransH	https://github.com/zqhead/TransH	Knowledge Graph Embedding by Translating on Hyperplanes 6
SimplE	https://github.com/baharefatemi/SimplE	SimplE Embedding for Link Prediction in Knowledge Graphs 7
RotatE	https://github.com/DeepGraphLearning/KnowledgeGraphEmbedding	RotatE: Knowledge Graph Embedding by Relational Rotation in Complex Space 8

Paper Knowledge Extraction

The pipeline of Paper Knowledge Extraction is in folder /code/paperknowledge/

• BERT-based QA (/code/paperknowledge/bertQA)

  • run single QA, with parameters: train_batch_size，learning_rate

  python run_squad.py --do_predict

  • run batch QA, with parameters: start file, end file, and year

  CUDA_VISIBLE_DEVICES=1 python run_test.py 116 125 2015

  • /code/paperknowledeg/bertQA/model can be download via google drive

• Paper Knowledge Entities Extraction (/code/paperknowledge/paperentity)

  • run

  python example.py

  • /code/paperknowledge/paperentity/dde and related files can be download via google drive

Citation

Our paper has been accepted as a resource paper at CIKM-2021 :

@inproceedings{deng2021gakg,
  title={GAKG: A Multimodal Geoscience Academic Knowledge Graph},
  author={Deng, Cheng and Jia, Yuting and Xu, Hui and Zhang, Chong and Tang, Jingyao and Fu, Luoyi and Zhang, Weinan and Zhang, Haisong and Wang, Xinbing and Zhou, Chenghu},
  booktitle={Proceedings of the 30th ACM International Conference on Information \& Knowledge Management},
  pages={4445--4454},
  year={2021}
}

System Introduction

We now introduce the GAKG platform: