By Alex Yeung.
Part 1: The Matching Process
Linking entities is a common theme in procurement data, but it is a significant problem in the UK. The concern with entity matching is the cost of false positives. This issue is particularly stark when finding the beneficiaries of money. The main challenge is that name matching is largely insufficient for people or companies, and here’s why.
Take for example, the “John Smith” conundrum.
“John Smith” – matches – “John Smith” is insufficient information to be useful.
“John Smith, Born: 1971/01/21, Reigate, Surrey” matches “Jon Smith, DoB: 21/01/1971, Surrey, UK”
This match is more useful because of the other pieces of data that corroborate this match, such as the date of birth or a string of text from the address. Even then for common names such as John Smith, it is possible that there might be two born on the same date who reside in the same town.
Lost In Translation
This problem is exacerbated by some names, especially those from the Far East, who are character-based. There’s often little consistency in how they’re represented in other languages. Take the author’s Chinese name: it can be anglicised to Saiman, Sai mun, or even Simon. The latter might not even be the native language name as many people adopt a name like ‘Thomas’.
Therefore, for John Smith and other names, much of the battle for matching is finding supporting data that allows the match to be made. This includes investigation using publicly available data to corroborate matches.
Of course, the reverse is also possible. Many cultures and countries have more unique naming systems such as longer and/or more unique names. Take the name of present UK chancellor Rishi Sunak. At the time of writing, on Companies House there are only three entries for company officers of this name, with two entries having the same month as a birth date. This narrows the scope for verification somewhat. Contrast this to 375,407 entries for a search for John Smith! (https://find-and-update.company-information.service.gov.uk/search/officers?q=john+smith#)
Batch Matching Names
Of course, when it comes to Companies House, a more efficient approach is needed. Even generously assuming that it would take 5 minutes on average to verify each name and 8-hour workdays, it would take over 3900 years to go through John Smith alone. To put that into context, the last known woolly mammoth reportedly died 4000 years ago. (https://www.sciencedirect.com/science/article/pii/S0277379119301398).
We Just Use A Script.
Using our data infrastructure and our algorithms, we can compare two lists within days, not millennia. Of course, we add our own investigatory magic to verify our matches. We have to. To give an example, here is what a machine might see:
Name list 1: glmb zimrh
Name list 2: glmb qznvh zimrh
Judging solely by what can be seen, a match is not immediately obvious in these two strings. In fact, what has been done is a simple inversion of the alphabet for the following two strings:
Tony James Arnis
Note: ‘Tony Arnis’ and ‘Tony James Arnis’ are both names made up for this article.
To the human eye, it’s pretty obvious that this is a match. This is because an Anglophone reader would have a priori knowledge of common surnames (Arnis is not a common surname: https://find-and-update.company-information.service.gov.uk/search/officers?q=arnis). The script does not have this a priori knowledge and would likely reject it. Of course, it can do: a more advanced algorithm trained on substantial amounts of prepared data might well address this but this takes time and resources to develop and train.
Stuck In The Middle
The ‘Tony James Arnis’ issue is representative of a broader challenge: middle names are really annoying for name matches because they are so inconsistently applied. A list of names might have middle names, it might not. Another list of names might have these names, it might not. Even personal use of middle names is not consistent. Incidentally, the UK passport only has so much space. If a person has too many middle names, one or more might be cut off partway. Again, this requires investigatory work to get right.
A Titular Distinction
Similar to the middle name problem is titles. Titles can appear anywhere within a name and means that script matching can reject what might appear to be perfectly good matches. Mrs, Dr., Prof., Professor, LLM, MCRVS, Eng, FBCS, these are but a few titles that can confound simpler scripts. This is especially troublesome when there are no consistent conventions for naming within any lists of names: some lists might put Dr. in the front of the name, others at the back. There are ways around this however through better scripts, but that’s a story for another time.
The challenges are many but not insurmountable. Matching is critical in bringing beneficial ownership data and procurement data together. With our colleagues at Open Ownership, we will continue to think about how we will do this in the future, both with the tools at our disposal and the tools we can create.