EUVIMEDEUVIMED entdecken
Crossref

Fine-grained Semantic Indexing of Biomedical Literature

Καντρίου, Μάριο Γκεζίμ

Aristotle University of Thessaloniki · 01.01.2023

DOI: 10.26262/heal.auth.ir.347233 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Aristotle University of Thessaloniki, erschienen am 01.01.2023. Verfasst von Καντρίου und 1 weiteren Autor:innen.

AbstractSchlagwörterBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Due to a considerable increase in the amount of data now available it has become increasingly difficult for people to locate pertinent information without the use of effective information retrieval strategies. To make it easier for researchers and practitioners to navigate biological material, medical subject headings (MeSH) were created. The process of adding relevant MeSH terms to biomedical publications can be time-consuming and expensive. Researchers are currently exploring the possibility of automating the annotation process using machine learning. Weakly supervised learning is good for controlling the assignment of new descriptors when there are not any annotated examples. The usefulness of contrastive learning and hierarchy techniques for fine-grained semantic indexing of biomedical literature was examined in this thesis along with the effects of balancing imbalanced data using weak labels and inserting scope notes and entry terms into the indexing process. It was found that weak label classification can be improved by using both contrastive learning and hierarchy. When these methodologies were used together the outcomes were noticeably better than when inadequate supervision was used exclusively. Additionally, adding scope notes improved the effectiveness of the outcomes. More in-depth information may be communicated by using longer language that included specific semantic elements rather than just difficult to manage words. The model with scope notes included was shown to be the most successful but the model with entry terms outperformed the baseline. The number of articles used in training had no effect on the F1-scores of the baseline model. Unoptimized models that relied only on hierarchy, contrastive learning or a mix of the two exhibited a positive association between the number of articles and their F1-scores. The addition of scope notes and entry terms to the optimized model revealed no correlation between the number of articles and F1-scores implying that adding extra articles would not greatly improve the performance of the models. The analysis of 18 labels revealed that a tool might be created to carry out fine-grained semantic indexing automatically. However, to support these findings more extensive research with more labels would need to be carried out.

Schlagwörter

semantic indexingbiomedical literaturemeshweakly supervised learningcontrastive learninghierarchical classificationscope notesclass imbalanceautomatic annotationinformation retrieval

Zitieren

APA

Καντρίου, & Γκεζίμ, Μ. (2023). Fine-grained Semantic Indexing of Biomedical Literature. Aristotle University of Thessaloniki. https://doi.org/10.26262/heal.auth.ir.347233

Vancouver

Καντρίου, Γκεζίμ Μ. Fine-grained Semantic Indexing of Biomedical Literature. Aristotle University of Thessaloniki. 2023. doi:10.26262/heal.auth.ir.347233

BibTeX
@article{2023,
  author = {Καντρίου and Γκεζίμ, Μάριο},
  title = {Fine-grained Semantic Indexing of Biomedical Literature},
  journal = {Aristotle University of Thessaloniki},
  year = {2023},
  doi = {10.26262/heal.auth.ir.347233}
}
Kostenlos starten →