Leveraging Text Mining in Insurance

By Karthik Balakrishnan, ISO Innovative Analytics

It is commonly believed that more than 90 percent of any organization's information is buried in unstructured data, such as text (documents, reports, notes) and pictures. This is particularly true in the property/casualty insurance industry, where unstructured data in the form of paper files, faxes, documents, adjuster notes, and underwriter comments abound. Recent advances in text mining technologies have made it possible to harvest valuable information buried in such unstructured data and leverage it for business results.Underwriting Gains Using Text Mining The business of insurance is all about risk assessment and management. To be successful, a carrier must identify, understand, and measure the true drivers of loss.

For instance, while claims systems often identify the type of Homeowners loss such as fire, water, wind, and hail, they do not pinpoint what caused the loss. It is hard to know, systematically, whether the water damage was caused by a burst pipe, faulty washer, overflowing bathtub, or rainwater through an open window.

Applying text mining to adjuster notes can reveal water losses caused by concepts such as "burst pipe," "washer," "bathtub," "kitchen sink," and "fish bowl." By building an appropriate classification of concepts, one can also separate water losses into weather-related versus non-weather-related.

With cause-of-loss tags extracted using text mining, many types of business outcomes can be conceived - from building underwriting and pricing models by type of peril to geographic characterization of loss distributions and trends. For instance, one might discover that frozen pipes drive water losses in Idaho, while malfunctioning washers drive water losses in California.

With such insights, carriers can develop better underwriting criteria to screen risks, while also establishing better loss control mechanisms, such as installing plastic overflow trays under washers of their customers.

Claims Analytics Using Text Mining Claims is a key business area for any carrier, with significant impact on bottom-line results. Much rides on the expertise and follow-through of the adjusters, who not only have to negotiate a fair and equitable settlement, but also make a number of other assessments such as subrogation, suspicion, and the need for an independent medical exam. Missed opportunities in making the right assessments can significantly impact business results.

Text mining can be used to uncover insights from adjuster notes and aid the systematic detection and referral of claims to such specialists.

For instance, text mining can extract fraud-related concepts such as "excessive treatment" from adjuster notes by searching for phrases such as "over treatment," "treatment appears exaggerated," "unnecessary treatment," and "buildup." Once built, the extracted concepts can function just as other structured data, since for each claim the concept "excessive treatment" will have a structured value of "yes" or "no" based on the text mining of the claim notes. This data field can be used with other data elements to create decision models for fraud referrals, as shown in Figure 1.

Figure 1: Model for Fraud Detection

Similarly, text mining can help identify subrogation concepts such as "other party at fault," "other party identified," and "insured not at fault" from adjuster notes. In this case, building a comprehensive dictionary of other parties (driver, repairman, roofer, etc.) can help the system determine if someone else was at fault. For instance, "furnace repairman incorrectly installed," "other driver struck our insured," "insured struck by adverse party," and "caused by roofer" would all trigger the concept of "other party at fault."

Once such concepts have been created, decision logic can be developed for systematic referral of cases with subrogation opportunity.

As these examples illustrate, text mining is a very relevant and useful capability that can be easily leveraged to create value across multiple functional areas of an insurance business. It can bring to light critical insights from textual notes of any kind - including adjuster, underwriter, loss control, auditor and customer representative notes.

About the Author: Karthik Balakrishnan, Ph.D., is vice president of analytics at ISO Innovative Analytics (IIA). A unit of ISO, IIA is focused on delivering advanced predictive analytics tools to the property/casualty insurance industry. He can be reached at [email protected].Recent advances in text mining technologies have made it possible to harvest valuable information buried in unstructured data and leverage it for business results.