RCMR 280k: Refined Corpus for Move Recognition Based on PubMed Abstracts

ABSTRACT Existing datasets for move recognition, such as PubMed 200k RCT, exhibit several problems that significantly impact recognition performance, especially for Background and Objective labels. In order to improve the move recognition performance, we introduce a method and construct a refined corpus based on PubMed, named RCMR 280k. This corpus comprises approximately 280,000 structured abstracts, totaling 3,386,008 sentences, each sentence is labeled with one of five categories: Background, Objective, Method, Result, or Conclusion. We also construct a subset of RCMR, named RCMR_RCT, corresponding to medical subdomain of RCTs. We conduct comparison experiments using our RCMR, RCMR_RCT with PubMed 380k and PubMed 200k RCT, respectively. The best results, obtained using the MSMBERT model, show that: (1) our RCMR outperforms PubMed 380k by 0.82%, while our RCMR_RCT outperforms PubMed 200k RCT by 9.35%; (2) compared with PubMed 380k, our corpus achieve better improvement on the Results and Conclusions categories, with average F1 performance improves 1% and 0.82%, respectively; (3) compared with PubMed 200k RCT, our corpus significantly improves the performance in the Background and Objective categories, with average F1 scores improves 28.31% and 37.22%, respectively. To the best of our knowledge, our RCMR is among the rarely high-quality, resource-rich refined PubMed corpora available. Our work in this paper has been applied in the SciAIEngine, which is openly accessible for researchers to conduct move recognition task.