Artificial Intelligence Initiative: European Central Bank

Athena, the European Central Bank’s (ECB’s) new natural language processing (NLP) artificial intelligence (AI) tool, is helping to address supervisory and regulatory blind spots. It will allow over 1,000 supervisors to analyse more than 5 million documents across the Single Supervisory Mechanism (SSM).

News articles, supervisory assessments and bank documents are fed into Athena, which allows users to find, extract and compare information in a single web-based platform, Steven Moons, team lead in the prudential domain services division at the ECB, tells Central Banking.

Supervisors working on site inspections have access to an English-language, machine-readable version of credit files that come in local languages.

The AI technology also allows officials performing horizontal analysis to efficiently check for compliance and consistency of regulatory standards, guidance and methodologies. When setting supervisory priorities, Athena allows officials to identify trends in the market and emergent risks.

“In the age of information, embedding AI in our day-to-day work is essential,” Daniela Schackis, deputy director general of SSM governance and operations, says. Developed with data analytics company Squirro, the ECB also set up a graphics processing unit-powered cluster to generate insights from the large language models, at scale.

The machine learning models are trained to assess document types, classify data on a hierarchy of topics, determine trending topics, perform sentiment analysis and identify references to supervised institutions via entity recognition.

Athena allows officials to “spot and assess new risks at the banking sector level and at an individual bank level”, Schackis says. Officials can look at annual review findings and measures raised by supervisors that a single bank needs to implement or follow up on. Outliers can be identified, and the evolution of banks tracked over time. “With Athena, we perform analyses quicker and more consistently, allowing us to be more proactive in our supervision,” she adds.

To build the NLP models, data scientists gathered labelled data covering a broad range of risk topics and used these to adjust the predictions to the supervisory context. “Though the machine learning models make suggestions and predictions, we recognise the value of expert judgement, which will continue to play an essential role in supervision,” says Ioana Karger, team lead in the suptech team at the ECB.

“We are committed to keeping the human in the loop,” she adds. “It is the supervisor who is able to go to the text, read the context, assess it and provide feedback on these enrichments, creating an iterative machine learning process informed by users’ expertise.”

Next, the introduction of a note-taking functionality will enable supervisors to annotate analysed content. Sharing these will improve transparency in the decision-making process as it becomes more auditable and consistent. The high-quality input data can also be used to advance AI functionalities.

The ECB is exchanging knowledge on the development and use of AI models with other institutions. “Beyond the SSM, we constantly exchange on developments in the field of NLP with our peers, such as the Bank of England, [US] Federal Reserve Board and the Central Bank of Brazil,” Schackis says. “We are really trying to align on the latest best practices and see how we can also collaborate further in this field.”

“Looking to the future, the potential for combining insights from structured and unstructured data is even more powerful,” Moons adds. “Being at the frontier of applying AI pushes us to continuously evaluate our practices.”