AB-XLNet: Named Entity Recognition Tool for Health Information Technology Standardization

Abstract

We conducted a study to identify drug-related information on non-standardized on non-standardized discharge summaries using a pre-trained BERT-based model. After tokenizing the dataset, it was identified with the IOB tagging schema and trained on the training data with the Random Insert technique through the pre-trained BERT. As a result, the F1-score of AB-XLNet was improved by 3% compared to XLNet, and ADE and Form, which could not be extracted from XLNet, were extracted. Future research will focus on presenting a generalized model using large amounts of data from multiple institutions.