EMoDi: Entity-Enhanced Momentum-Difference Contrastive Learning for Semantic-Aware Verification of Scientific Information
Ze Yang, Yimeng Sun, Takao Nakaguchi, Masaharu Imai · 2023
This paper proposes the EMoDi system to improve the performance of the entire scientific information verification pipeline. First, the Momentum-Difference contrastive learning framework is introduced to capture more semantics information. In abstract retrieval, entity-enhancement and noise-ignoration are introduced to improve the ability to retrieve relevant abstracts more accurately. In addition, a two-step verification method is used in label prediction to improve the label prediction ability and reduce the false positive rate of the “NOT ENOUGH INFO” label. The proposed pipeline outperforms the baseline VERISCI and QMUL-SDS. The code of this system is available on GitHub.