Claim processing in insurance is at the core of its operations and business. Even in the case when claims are processed entirely electronically it is rather hard to assess new claims. The industry standard is to use Rule Engines with a large set of expert-written rules. The limitation to this approach is increasing complexity of rules and the related maintenance. Additionally these systems fail to classify a significant portion of incoming claims.
In this talk we describe the automation of the processing for complex claims using machine learning methods: We present the journey we took together with our client starting with the initial proof of concept, through the definition of the business case and the implementation of the machine learning algorithm, all the way to the deployment and the operationalisation of the new system.
In our talk we particularly focus on how we used state of the art NLP methods, such as word2vec and sent2vec, to define a similarity measure and learn a sparse space of more than 100’000 features. Additionally we describe how we worked together with end users to build an intuitive validation framework, thus building trust in the approach.