Exploring author name disambiguation on PubMed-scale

Author name disambiguation (AND) creates a daunting challenge in that disambiguation techniques often draw false conclusions when applied to incomplete or incorrect publication data. It becomes a more critical issue in the biomedical domain where PubMed articles are written by a wide range of researchers internationally. To tackle this issue, we create a carefully hand-crafted training set drawn from the entire PubMed collection by going through multiple iterations. We assess the quality of our training set by comparing it with SCOPUS-based training set. In addition, for the performance enhancement of the AND techniques, we propose a new set of publication features extracted by text mining techniques. The results of the experiments show that all four supervised learning techniques (Random Forest, C4.5, KNN, and SVM) with the new publication features (called NER model) achieve improved performance over the baseline and hybrid edit distance model.