AI Case Study
The Bank of England improves its supervisory capabilities by automatically identifying potential regulatory issues from public documents published by over 500 insurance firms using natural language processing text classification
Bank of England uses Cognition, an AI solution that uses natural language processing to mine text from documents published by 100s of insurance firms it is supervising, to train new models to assess regulatory and business risk with input from trained analysts. The platform uses supervised learning techniques and natural language processing.
"The Bank of England uses Cognition, a machine education tool built by Digital Reasoning to train artificial intelligence models.
The Bank found the Cognition “interface to be intuitive, which enabled analysts to train models quickly with little formal training.” It was “successfully able to use the model to categorise instances of business risk and external context in the data.”
Cognition enables organizations to transfer their employees’ knowledge to computers, quickly and accurately. Organizations that achieve this are best placed to realize the competitive advantages of automation and insights at scale. Cognition uses active learning and intuitive interfaces that empower anyone to train AI models. It cuts training time from months to days or even hours, demands no data science skills, and reduces the costs of analytics model building by orders of magnitude."
Legal And Compliance
"The Bank’s Prudential Regulation Authority (PRA) is the United Kingdom’s prudential regulator of deposit-takers, insurers and major investment firms. As part of its remit, the PRA supervises over 500 insurance firms including general insurers, life insurers, friendly societies and the London insurance market.
Bank of England was exploring new technology to automate extracting policy-relevant information from the weekly texts published by 100s of firms."
Training time reduced to days from months.
More efficient analytical model.
"Approximately 10,000 data points were loaded into the tool in order to train it with input from experienced analysts.
The Cognition tool enabled users to build ‘annotation models’ based on particular themes or ‘labels’. For each model, users applied labels to excerpts of text selected by the software. In this case, models were trained in line with the PRA’s supervisory risk framework, which considers the overall risk context for insurance firms in terms of ‘external context’ and ‘business risk.
During the training process, Cognition used supervised machine learning techniques to adjust the model iteratively in order to improve understanding of which sections of text might be considered as examples of business risk and external context respectively. The tool also features a ‘cross-validation’ feature, which provides information on the accuracy of each model as it is trained by the user using an increasing number of examples.
The tool also included a multi-labelling feature, allowing users to train multiple models at the same time (for example to identify text relating to both business risk and external context within the same
data set). We found that this enabled training to be carried out more efficiently."
Trained the model using 10,000 data points. Public documents published by over 500 insurance firms.