Exploring Transfer Learning for Insurance Applications

Prior to the development of OpenAI's Large Language Model, which provided a striking illustration of how text parsing and generation could be handled using AI instead of human resources, this experiment was conducted. I recollect that its original form was quite like a research paper, requiring some reworking and reshaping to make it more easy to grasp. Ultimately, the trial went no further than the experimental phase; I presume this is because of a cautionary attitude towards the degree of data integration it would have entailed. This experiment is designed to investigate if the Transfer Learning machine learning algorithm can further enhance the summarization and categorization of incoming insurance policy inquiries. Transfer Learning is a type of Deep Learning algorithm that utilizes previous knowledge from related tasks to assist in solving current, related questions. The FI saw that Transfer Learning could be utilized because there were multitude of use-cases in which machine learning algorithms couldn't be deployed due to a lack of training data specifically related to the problem at hand. For instance, in regards to the task of processing emails pertaining to insurance policies, there may be a shortage of data but there is a wealth of data connected to customer inquiries about various insurance policy elements. The FI is exploring the possibility of employing a machine learning algorithm to generate and condense text related to insurance policies. This would enable the summarization of client emails seeking to purchase insurance, and ensure that they are directed to the right contact. the process of exchanging documentation is complicated and laborious, and the technology behind it has not kept up. This use-case was chosen because the commercial insurance industry is highly manual, with widespread use of emails, MS Word, and PDFs to communicate. The difficulty and laboriousness of exchanging documentation, combined with outdated technology, are both contributing factors. Commercial insurers depend on sizeable teams to manage incoming emails, MS Word documents, and PDFs. Their most profitable emails involve new insurance or renewals from current customers. Since many insurers receive quotes from multiple providers, a lag in response time could mean forfeited business. Therefore, commercial insurers must respond promptly to new and renewal insurance in order to remain competitive and are accordingly compelled to maintain a substantial email processing team. The FI intended to enhance the handling of incoming emails by utilizing a trained machine learning algorithm for the purpose of enhancing the summarization and categorization of such emails. If successful, the POC could open up a new possibility for insurers to deal with incoming emails. At present, their options to enhance the productivity of their email processing team are restricted to (1) giving more training, though there is a limit to how efficient someone can be; (2) adding personnel, which can be costly; or (3) outsourcing, which comes with its own difficulties in terms of quality control. Transfer Learning might, potentially, offer a technological alternative, potentially offering the email processing team increased efficiency.

Source: https://medium.com/@notherArmchairEconomist/experiment-with-transfer-learning-on-insurance-use-case-e6da35ad9fdd?source=rss------fintech-5