Finra taps AWS for next-gen regulatory search tools

The Financial Industry Regulatory Authority (Finra) plans to revamp its enterprise search capabilities using tools from Amazon Web Services to make it easier for its internal examinations staff and third parties, such as brokers and investors, to find information in its filings and reports using natural-language search to generate more relevant results.

Finra’s goal is to build “regulatory connected intelligence by leveraging artificial intelligence and machine learning,” that will not only return lists of documents as search results, but will also provide links between entities in the results, a “factory” written summary of the results, and analytics that deliver personalized results based on a user and what organizations, individuals, stock symbols, or other criteria they are searching for, said Dmytro Dolgopolov, senior director at Finra, during a recent AWS webcast.

Enterprise search is not a new concept for Finra. It built its first enterprise search platform in 2009, providing a single place to search all the structured and unstructured information collected by the regulator.

“Today we have dozens of sources and millions of structured and unstructured records. And every day, more than 70% of staff at Finra use this enterprise search,” Dolgopolov said. “But any solution, over time, can encounter challenges.”

Finra identified several key challenges: its existing platform used keyword search but did not support more complex queries or natural language search; it lacked personalization insofar as it did not return results specific to users based on their role and current assignments, and it did not integrate with Finra’s knowledge graph to improve the relevancy of results.

Finra—which in 2014 moved 90% of its data to AWS’ cloud to process and store billions of incoming transaction records daily, and also uses a range of Amazon tools to perform billions of validation checks per day as part of its Oats (Order Audit Trail System)—decided to investigate how the tech giant could help it achieve these goals.

The new initiative focuses on using natural-language search to extract information from unstructured documents using AWS’ Textract tool and its Neptune graph database, to identify relationships between data points using AWS Comprehend, and to query the information using Amazon Kendra, an intelligent search service that responds to natural-language questions using machine learning to search unstructured data.

Rather than relying on simple keyword searches, Kendra is able to understand questions written in plain text, and can deduce the “intent” of a question, regardless of how a user phrases their query. AWS pre-trains Kendra on domain-specific terminology, including for financial services, and performs ongoing “relevance tuning.”

Proving ground

Daniel O’Brien, director of Finra’s enterprise analytics program, speaking on the same webcast, says the regulator first chose to test out the tools on its rulebook—a mix of structured, semi-structured, and unstructured information that is frequently referenced by Finra staff.

“To do this, we first crawled the rulebook pages. We broke down sections into sub-sections, extracted the associated metadata, put them into an S3 (AWS’ Simple Storage Service cloud) bucket, set up the Kendra index, created all the necessary custom attributes for our use case, and repeated those steps for the rulebook FAQ section,” O’Brien said.

Not only did the search prove successful in returning relevant search results; it only took a few days to complete the process, O’Brien said.

Finra then tested a second scenario, using the combination of tools to search arbitration awards data on the Finra.org website.

“We wanted to ask a different set of questions of the data. These scenarios are typical of the types of case data Finra sees internally and is a good proof case for how Kendra could help internal stakeholders get information quickly,” O’Brien said. “We chose this dataset because it is less interpretive than the rulebook, offering more discrete answers.”

Finra tested it using a single question—to identify the claimant in a particular arbitration case—posed using natural language. O’Brien said Kendra returned the specific answer, whereas other search technologies returned highlighted sections of content. He also says the team tried asking the question many different ways, and no matter how poorly worded the question, Kendra was still able to interpret and understand it.

“Enhanced search capabilities using natural language understanding can reduce time-to-value for our examiners,” he said. “Users want answers, not more data … and factory responses reduce search-to-answer timelines.”

Finally, Finra connected its own knowledge graph to Kendra and Neptune to evaluate how this could improve the search experience.

“Our vision is to graphically be able to navigate relationships in the data,” O’Brien said. The combination was able to enrich the results with “associated dimensions” of related data. “This opens a world of opportunity to Finra, including similarity search, relevance management, and recommendations. The enriched relevance tuning with the additional dimensions of data are a high value-add to Finra overall with how we choose to rank our results in the future. In addition, leveraging that firm knowledge outside individual documents or sources of information creates a more complete user experience.”

Finra officials did not immediately respond to inquiries about the timeline for rolling out the new search capabilities.