A lot has been written about “phonetic algorithms” since Soundex was created for the US Census in (we think) 1880, but the world seemed to stand fairly still until computer software started to implement name matching in the 70’s. The strange thing is that Soundex seems to have remained the de facto standard until well into the 90’s, even up to the present day – strange because Soundex is manifestly prone to false matches as well as missing some typical true name matches. To explain why this happens, let’s look at how Soundex works:
Soundex constructs a crude non-phonetic key by keeping the initial letter of the name, then removing all vowels, plus the letters H, W and Y, and translating the remaining letters to numbers e.g. Tootill and Toothill both become T340. It gives the same number to letters that can be confused e.g. ‘m’ and ‘n’ both become 5. It also drops repeated consonants and consecutive letters that give the same number e.g. S and C. To illustrate some of the issues: Soundex translates Brady, Beard and Broad all to B630, and Wilkins and Wilson both to W425, and yet doesn’t match Knot and Nott – let alone more challenging examples like Dayton and Deighton.
A lot of the work done in the 70’s and 80’s focused on shoring up Soundex’s more obvious symptoms, rather than addressing the root problem – it doesn’t attempt to understand how people pronounce names. In the last 20 years, there have been various “phonetic” algorithms enter the public domain but Metaphone 3 is the only one that I know of that doesn’t invariably disregard the vowels after the first letter of the name.
Much of the material we read back in 1995 (when searching for a phonetic algorithm that worked better than Soundex) started off on the wrong track by adopting a similar approach to Soundex. Often, the authors quoted various sounds to support their solutions which are rarely present in names of people e.g. the different “ua” sounds in “persuade” and “casual”, the “pt” in “Ptolemy”. Even when I’ve revisited the subject in the intervals since then, there has been little advance. Back in 1995, we decided that 360Science would write our own genuinely phonetic algorithm, and we laid down these requirements:
- It must produce phonetic codes that represent typical pronunciations
- It should focus on “proper names” and not consider other words
- It should be loose enough to allow for British and American regional differences in pronunciation (e.g. “Shaw” and “Shore”) but not so loose as to equate names that sound completely different
- It should not try and address other forms of fuzzy matching that arise from keying or reading errors and inconsistencies – phonetic matching should, by definition, simply address words that may sound the same.
The last point is important – the most effective way to match names is to use a combination of techniques to allow for different kinds of error, rather than try to create one algorithm that is a jack of all trades. I will describe our phonetic algorithm in a later post, but software development, like data quality, is a journey and not a destination – so we’re always looking to improve.