Co-attention-Based Pairwise Learning for Author Name Disambiguation

Digital libraries face a pressing issue of author name ambiguity. This paper proposes a novel pairwise learning model for author name disambiguation, utilizing self-attention and co-attention mechanisms. The model integrates textual, discrete, and co-author attributes, amongst others, to capture comprehensive information from bibliographic records. It incorporates an optional random projection-based dimension reduction technique for efficiency to handle large datasets. The attention weight visualizations provide explanations for the model's predictions. Our experiments on a substantial bibliographic catalogue repository validate the model's effectiveness using accuracy, F1, and ROC AUC scores.