Séminaire du CIMMUL – Amedeo Zito

Loss prediction in insurance using NLP text representation methods

Amedeo Zito

Intact Financial Corporation
École d’actuariat, Université Laval

Résumé

Loss prediction in insurance has always traditionally used structured data as predictors. However, now with the growing access to state-of-the-art natural language processing models, it has become much easier to use textual data to enhance loss predictions. Moreover, the narrative content of an incident description may provide sufficient detail to assess its severity. In this essay, we demonstrate a framework on different approaches to transform textual data into vector representations that can directly be used as predictors for loss regression models. This is a very recent topic in actuarial science and therefore only little literature is available on that subject. Using the presented frameworks does not require any prior knowledge of the text. We present a variety of text representation methods used in NLP for several decades. In addition, we present state-of-the-art language models that encode valuable information into numerical vectors and show how they can be utilized in regression tasks. We also explore how the length of text can impact all the different approaches. Finally, we discuss methods like SHAP values and soft-attention to interpret and understand the model predictions.

Le séminaire aura lieu au local 3870 du pavillon Alexandre-Vachon et en ligne.

Du café sera servi. Emportez votre tasse !