Understanding Icelandic insurance claims : from text to incident categories

Abstract

In the field of insurance claims, there is much room for innovation and streamlining. Not only when working with the Icelandic language but also in terms of automating claim handling workflows. This thesis explores the possibility of understanding the context of a written incident report using supervised learning methods and how unsupervised clustering can be used to discover new incident categories. We use proprietary claim categories to label the incident reports and evaluate models based on their ability to perform the same classification. Transformer architectures, the state-of-the-art for Natural Language Processing tasks like this, show remarkable strengths in learning the context of incident reports. Surprisingly, traditional classifiers, such as Support Vector Machines, are showing similar strengths. By fine-tuning IceBERT, a large language model based on the Transformer architecture, we achieve a high accuracy of 93.2\%. This may be considered a perfect classification, considering that the initial incident reports only represent the first step in the claim process, with subsequent information potentially influencing the final incident category.