Framework of BERT-Based NLP Models for Frequency and Severity in Insurance Claims

It is challenging to incorporate textual information from insurance datasets for predictive modeling. We propose a framework for claim frequency and loss severity modeling based on a new natural language processing (NLP) technique, named BERT to extract textual descriptive information from claim records. Predictions are obtained using artificial neural networks (NN) for regression. Additionally, the shape of the predictive distribution is estimated and outlier treatment with corresponding data analysis is discussed. This research shows that BERT-based NN model provides a great possibility to outperform other models without using textual information in accuracy and stability when suitable textual data are available for modeling. This research outlines an automated procedure of BERT-based frequency-severity predictions for insurance claims. Address for Correspondence: sx2b@mtmail.mtsu.edu