SRL NPL for Sanctions Compliance
Here’s a simple question:
Why don’t watchlist screening vendors deploy semantic role labeling (SRL) methodologies when screening free-text fields?
The value proposition is obvious — analysts would save hours of work each week by not having to clear false-positive alerts like the match of the word “shipped” to “XHAFERI, Shefit” (SDN).
But seriously, this is a huge problem. If a large portion of a financial institution’s business activity is in Letters of Credit or similar transactions that involve sending large free-text messages blocks between banks, then these garbage matches can easily balloon to more than 50% of generated alert activity and hamper cost-center efficiency.
As mentioned in a previous post (here), two criteria must be satisfied for quality matches to be generated:
- The internal company data and sanctions record data must be comparable.
- The method of comparison must be appropriate for the type of records being compared.
Failing to implement SRL methodologies across free-text fields flies in the face of the second criteria.
Essentially, SRL breaks down a sentence into it’s principle components (think subject, object, action, etc.), like “Mary shipped little lamb” to:
- Subject: “MARY”
- Action: “SHIPPED”
- Object: “LITTLE LAMB”
Under this processing methodology, it wouldn’t make any sense to compare “shipped” to sanctions list or watchlist records – but, that’s what we do now with our current carpet-bomb string matching approach.
So again, why don’t watchlist screening vendors deploy SRL methodologies when screening free-text fields?