Publication Type Tagging using Transformer Models and Multi-Label Classification

Abstract

Indexing articles by their publication type and study design is essential for efficient search and filtering of the biomedical literature, but is understudied compared to indexing by MeSH topical terms. In this study, we leveraged the human-curated publication types and study designs in PubMed to generate a dataset of more than 1.2M articles (titles and abstracts) and used state-of-the-art Transformer-based models for automatic tagging of publication types and study designs. Specifically, we trained PubMedBERT-based models using a multi-label classification approach, and explored undersampling, feature verbalization, and contrastive learning to improve model performance. Our results show that PubMedBERT provides a strong baseline for publication type and study design indexing; undersampling, feature verbalization, and unsupervised constrastive loss have a positive impact on performance, whereas supervised contrastive learning degrades the performance. We obtained the best overall performance with 80% undersampling and feature verbalization (0.632 macro-F₁, 0.969 macro-AUC). The model outperformed previous models (MultiTagger) across all metrics and the performance difference was statistically significant (p < 0.001). Despite its stronger performance, the model still has room for improvement and future work could explore features based on full-text as well as model interpretability. We make our data and code available at https://github.com/ScienceNLP-Lab/MultiTagger-v2/tree/main/AMIA.

Figure 1:

Flow diagram of experiments including data undersampling, feature augmentation, and contrastive learning. The dense layer used for contrastive learning experiments and the linear layer used for label predictions utilize the [CLS] token’s embedding from the last hidden state layer within PubMedBERT.

Figure 2:

The sub-figure on the left shows the PT label distribution for all articles in the dataset. The right sub-figure shows the individual label performances using the best-performing model (80% undersampling and verbalized feature augmentation) on the test set.