Proceedings of the 3rd International Conference on Bigdata Blockchain and Economy Management, ICBBEM 2024, March 29–31, 2024, Wuhan, China

Research Article

Building High-quality Psychology Knowledge Graphs from Text using REBEL

Download12 downloads
  • @INPROCEEDINGS{10.4108/eai.29-3-2024.2347315,
        author={Zeli  Chen and Yixuan  Chen and Raymond  Xu and Yao  Zhao},
        title={Building High-quality Psychology Knowledge Graphs from Text using REBEL},
        proceedings={Proceedings of the 3rd International Conference on Bigdata Blockchain and Economy Management, ICBBEM 2024, March 29--31, 2024, Wuhan, China},
        publisher={EAI},
        proceedings_a={ICBBEM},
        year={2024},
        month={6},
        keywords={nlp (natural language processing) text-to-knowledge graph rebel relation extraction psychology},
        doi={10.4108/eai.29-3-2024.2347315}
    }
    
  • Zeli Chen
    Yixuan Chen
    Raymond Xu
    Yao Zhao
    Year: 2024
    Building High-quality Psychology Knowledge Graphs from Text using REBEL
    ICBBEM
    EAI
    DOI: 10.4108/eai.29-3-2024.2347315
Zeli Chen1,*, Yixuan Chen2, Raymond Xu3, Yao Zhao4
  • 1: Lafayette College
  • 2: The Experimental High School Attached to Beijing Normal University
  • 3: Wayland High School
  • 4: Guangzhou City University of Technology
*Contact email: chenzel@lafayette.edu

Abstract

This paper delves into the progression of Knowledge Graphs since their inception in 1972, highlighting their applications in sectors such as healthcare and finance. We present a novel model that automates knowledge graph creation using NLP technologies like BERT, spaCy, and NLTK. The study also explores three key tool categories in knowledge graph construction: Linguistic, Autoregressive Encoder-Decoder, and BERT-based REBEL, assessing their effectiveness and adaptability. Utilizing advanced algorithms like REBEL and KG-BERT, our system expedites the training process and enhances the quality of Knowledge Graphs. Testing on 20 'Psychology'-focused Wikipedia articles, we achieve near-optimal results in just 8 epochs. The study concludes that pre-training and predictive language models hold significant promise for improving knowledge graph utility and construction.