Adverse Media Screening Without Multilingual NLP: Why English-Only Coverage Is a Compliance Liability
A compliance analyst at a European bank checks a new corporate client. There are no sanctions list hits, PEP list hits, or watch list hits found with the name. The adverse media check only returns one article, in English and from a neutral trade publication. The customer successfully completes the onboarding process.
A regional prosecutor in Southeast Asia later charges the same man with money laundering, 18 months later. It had all been investigated in the press in the local language for more than a year, but it was not written in English and the bank’s adverse media screening software did not deal in English-language sources.
This is a known failure mode in AML compliance, in different forms. The quality of adverse media screening based on information available in the English language is systematically lower in jurisdictions where the most operationally relevant financial crime reporting is otherwise in local language. This is no hypothetical problem for financial institutions with international exposure; it’s a compliance liability.
What Adverse Media Screening Really needs to do
Adverse media screening is a part of Customer Due Diligence (CDD) and Enhanced Due Diligence (EDD) as outlined in the FATF Recommendations, the UK Money Laundering Regulations 2017 and the EU Anti-Money Laundering Directives. The aim is to detect negative data about customers, beneficial owners and the counterparties that is not found in structured data sources such as sanctions lists, PEP registers, court data, etc. but does exist in public media data sources.
The Wolfsberg Group’s AML Principles refer to adverse media screening as an essential part of the CDD process where the financial institution should search for information on financial crime, corruption, money laundering and financial sanctions. Not only do they have to be domestic sources, but they must not be limited to English sources.
In their enforcement activity – which has included examination of adverse media failures in a number of recent actions – regulators consider whether an institution has a process in place, whether that process is proportionate to the risk and whether the coverage is adequate to identify publicly available negative information regarding the customer. One of those sufficiency analyses is language coverage.
Only Adverse Media Coverage Fails
Financial crime reporting in emerging markets, high risk jurisdictions and non-OECD countries is mainly available in local language. These investigations into corruption in Latin America are covered in Spanish and Portuguese. Prosecutions for fraud are reported in Southeast Asia in Indonesian, Thai, Vietnamese and Tagalog. There are Arabic documents which document sanctions enforcement in the Gulf. Eastern European financial crimes are committed in the native language, Russian, Polish and Romanian.
If the case is big enough to warrant an international report, a compliance tool that accepts only English-language sources will be able to capture the wire service version of the story of a major case. It will fail to capture the investigative reporting into money-lenders that happens in the region, court filings reported by the local press, regulatory enforcement action notices and the coverage of financial crime issues in the business media, which form the main source of the record of financial crime in those markets.
The gap is a function of the financial institutions’ exposure to the jurisdictions where they onboard customers and/or have business activity. As they trade with non-English-language counterparties, the weaker their adverse media program is.
The False Positive Problem in Adverse Media
If adverse media screening is not coupled with intelligent filtering, it will produce a lot of false positives. Each of these can generate a signal, which needs to be manually checked: Common name matches, geographic associations, and tangential mentions of a customer’s name in negative news. With a non-AI based adverse media triage, the percentage of adverse media alerts in the total number of alerts can be a large percentage — as well as a disproportionately large percentage of analysts’ time spent on irrelevant results.
There are two aspects of the false positive issue in adverse media. Source relevance is the first; if a name is mentioned in a news article that has nothing to do with financial crime, then this should not trigger a reviewable alert. The second is sentiment accuracy, defined as a neutral factual reference to a customer in a story that is focused on someone else’s fraudulent activity is quite distinct from stories that directly reference the customer’s own wrongdoing.
If keyword proximity is the only criteria used, and there is no sentiment analysis or risk categorization based on NLP, adverse media tools generate alerts based on proximity alone. The analyst queue becomes filled with irrelevant hits, actual risks are lost in the noise and the screening process doesn’t offer as much protection as the number of alerts suggests.
How Multilingual NLP-Powered Adverse Media Screening is Different
Natural language processing (NLP) applied to adverse media screening alters the capability in two essential methods. The first one is source coverage: with multilingual corpora for training, an NLP model can process and classify news articles in any language, including non-English, meaning that it can provide global coverage without having to be manually translated.
The second one is risk classification, instead of marking any article that has a customer’s name and a keyword selected from a list of risk terms, NLP models are used to analyse the semantic relation between the customer and the reported negative event. If the article uses the word “customer” as opposed to the word “suspect”, it is classified differently. This article is not about financial crime exposure at the customer’s own financial crime exposure, but instead mentions financial crime in a country.
Intelligent risk tagging allows the categorization of adverse media hits by type – financial crime, corruption, terrorism financing, regulatory action, fraud, reputational risk – to which a risk-based review priority can be given to the compliance team. A regulatory enforcement action against a customer’s business is a risk item that is different than a reputational risk item. Both should be detected.
AML Watcher’s solution to navigating the multilingual adverse media challenge
AML Watcher’s adverse media screening capability consumes 5,000+ sources from around the world in multiple languages and in countries. Intelligent risk tagging across the categories that are relevant to AML compliance – Financial crime, fraud, corruption, terrorism financing, sanctions violations, regulatory action and reputational risk – NLP powered risk detection processes sources, irrespective of language.
Sentiment analysis helps differentiate between a negative factual mention of a customer versus a neutral mention, eliminating one of the biggest sources of false positive volume. The TruRisk AI agent filters out non-risk-relevant adverse media hits and provides results to analysts prioritized and categorized by risk signals.
The AML Watcher adverse media screening capability gives compliance teams the language coverage and analytical intelligence for which regulators are looking, and offers them a way to assess whether their adverse media screening program is proportionate to their geographic and customer risk exposure.
Steps to Creation of a Strong Adverse Media Screening Program: Considerations
Identify the languages of your customer base and counterparty jurisdiction, represents your language coverage minimum requirement.
Review the list of sources for your existing adverse media tool: How many sources, countries, language?
Check if your tool uses NLP sentiment analysis or keyword proximity matching.
Analyze your taxonomy of risks: are there categories that are in line with Wolfsberg and FATFs definition of relevant categories?
Set alert thresholds differently for high-risk customers, lower thresholds for low-risk customers.
If you have a program of adverse media, record the rationale for your program in your AML risk assessment as evidence of a proportionate approach.
Adverse media screening which cannot understand the language in which the most relevant financial crime reporting is conducted is no minor mistake — it’s a real blind spot. It demands not only coverage breadth but analytical sophistication as well to fix it. Both need to cooperate with one another.
