Cross-Script ID Matching & Screening: Reducing False Positives & Demystifying Explainable AI

Declan Trezise, Global Head of Consulting at Babel Street demystified the latest ways and means of addressing KYC and AML compliance issues for guests at the Hubbis Digital Wealth Forum in Singapore on October 4. He explained the impact of false positives and missed matches in screening and filtering. He told delegates how AI-based smart fuzzy name-matching can work. And he then talked them through matches for regulatory approval and the ongoing drive towards continuous improvement in such protocols. Babel Street is, he stated, at the cutting edge of technology solutions for state-of-the-art onboarding, KYC and AML challenges the wealth management industry faces, using explainable algorithms that also span into the worlds of counter-terrorism and border control in which the firm is also a leader.

Babel Street presents itself as a new force at the cutting edge of countering the volume, velocity, and variety of identity, onboarding, KYC and AML threats and risks, which the firm says are constantly multiplying. “Babel Street provides the most extensive, relevant reach combined with AI and natural language processing to empower our customers to stay ahead of these dangers,” Declan explains. “We are here speaking today alongside our partner, RegTech group IMTF, and to explain the role we play in helping fight financial crime.”

Declan also introduced himself, as somewhat – in his own words - of a “nerd” by inclination and nature, but one who has extensive experience working with parties to help control risks in the world of financial, counter-terrorism and border control, all areas that Babel addresses and is a leader in. “I am also the type of person who before getting into investing in cryptocurrencies had to write my own algorithm so I could try to decipher that space, so, yes, I really am that person,” he quipped.

Solving real world KYC issues

He told delegates that AI solutions can nowadays be increasingly deployed to improve a lot of the onboarding, KYC and AML processes that wealth management practitioners will already be familiar with. “And specifically, I'm talking about the tactical use of smaller language models rather than large language models,” he added. “These help significantly in terms of that ongoing perpetual KYC requirement, to improve the efficiency around onboarding and screening customers, and also in terms of sanction checking.”

He stated that if you can automate as much of the process as possible, and if you can take the humans out of the loop and be entirely confident in the algorithms applied in their place, you can take your onboarding from months, weeks, or days down to something that could take hours or minutes. “We integrate with platforms such as IMTF to create and make transparent the automated workflows to help their clients gain customers and retain the customers and weed out the dangers.”

Unmasking the risks

Declan said this automation covers standard protocols such as analysing and then verifying biographical information and documents, checking against watch lists, and all those related checks. “But you can go further,” he said, “by checking if that name is in the news using a sophisticated and intelligent, smart body matching engine. This means understanding the connections between mentions of identities in the news and aligning them specifically to risk detection as the next stage.”

He also pointed to transaction screening, which he said needs to happen incredibly quickly. “Our top four customers make on average 700 million watchlist checks nightly using our algorithm,” he reported.  “And we have extensive reach if you entered the United States in the last six years, it was extremely likely it was our name-matching algorithm that checked your identity.”

AI-enabled

He explained their AI fuelled, intelligent or smart fuzzy name matching capability product goes by the name of Rosette. “It's able to take a name written in one script, in one language, and with parts, missing parts, added parts, transposed and mixed up, spelt incorrectly with pronouns, with titles, with suffixes, and can then always find the correct match for that name,” Declan elucidated. But the algorithm would be useless if we could not have it as explainable to an end user or a regulator, but this product is exactly that – understandable to those parties.”

Delivering in partnership

Moreover, he said that working with partners such as IMTF, they can actually tailor and adapt our engine and its parameters to become much more precisely useful and appropriate for the wealth market customers IMTF or others work with. “An algorithm is great, better if it's explainable, but even better if it is adaptable and flexible,” he added.

He highlighted the recent uncovering of a vast money laundering operation that had been active for years. “Would your existing systems have detected these individuals?” he pondered. “The answer is for the most part, ‘no’, and then we can think about what was missing, because inevitably all these people have real identities that were verified and checked against the existing systems, but not found out until years later. But it is the tip of the iceberg and not related only to finance, but also to terrorism and other crime.”

Staying ahead of the game

He explained that their solutions are regulator-approved and they work constantly with the authorities to stay ahead of needs and challenges. He reported they are also class-leading in terms of accuracy, employing an integrated approach to name screening and search to massively reduce false positives using the intelligent algorithm.

“And the big one is maximising the F1 score, which is a metric that measures the accuracy of a search and match system,” he explained, offering delegates considerably more detail on their approach and technicalities around name matching, translation, recognition and so forth.

Do you speak every language?

“For a truly global screening system, you need to understand names from all countries in all different languages and even scripts, and the ways those might inter-relate or connect,” he explained. “And that is what we can do – we are able to continually add languages to our screening capability and our own algorithms.”

“Recently we have been adding Khmer, Malay, Indonesian, Vietnamese, and others in this region alone.”

Can we understand? Yes, we can

He closed the talk by offering the guests some further insight into the solution’s explainability. “If the algorithm is deterministic, it will always give you the same score every time you run it, but instead of a yes or no answer, you can also obtain a graded answer, a score between 0 and 100.  You can then set a threshold for the match, not using some number arbitrarily, but decided by data, and that gives you ultimately the explainability you need in a regulated environment.”