This project was created as a coursework submission for our Interactive AI Group Project.
Based on the requirements of L.V Insurance, our group developed a backend question-answering system which assists the service desk staff in extracting relevant information from customer phone calls. The Passive Information Extraction System (P.I.E.S) analyses a live transcript of the conversation to improve customer experience by allowing service desk operators to concentrate on the human interaction rather than data collection. The system processes all information from the conversation in real time and enters it on the system while the call operator concentrates on the customer’s welfare. All information extracted can be verified by the call handler ensuring vital information is never missed or incorrectly identified. The system uses a BERT model trained on the Stanford Question Answering Dataset (SQuAD). One of the main advantages of the implemented system is the ability to produce more training data with every call the company handles. With thousands of calls per week, it is possible to generate a sufficiently large labelled dataset of transcripts specific to the company’s requirements. This allows for routine updating of the Question and Answer model to ensure it is performing to a high standard.