Content Copyright © 2006 Bloor. All Rights Reserved.
Long time readers of this column will know that I espouse the cause of semantic approaches in a variety of areas: I think that natural language processing works better in search engines, and I think that LAS (which has just been acquired by IBM) offers just about the best name matching on the market, thanks to its semantic and linguistic basis as opposed to the statistical approach that is common amongst data quality vendors.
As it happens, I have recently uncovered another supplier in the data quality space that espouses the cause of semantics as opposed to statistics, which is Silver Creek Systems. However, in this case, the company is not focused on name matching but on product matching.
Product matching is an order of magnitude more complex than conventional name and/or address matching. For example, suppose that you sell electrical resistors. Now, resistors have a variety of attributes: resistance, power, tolerance, general description, manufacturer, UNSPCS and FSC industry codes, and so on. Now, you get these product details in from your suppliers: how do you build a consistent catalogue when you bear in mind that these details may come in any order, that you have numeric values mixed in with letters, and that you may have electrical symbols (Ω instead of ohms, for example)?
Traditional approaches to this sort of matching use a statistical methodology: that is, you look for patterns within the data. The problem is that with product and similarly complex data derived from diverse sources, the data is all intermingled and it is difficult to extract relevant patterns. While there are some relatively simple product-based environments in which traditional methods can work well, in more complex situations involving such things as electrical components, electronic consumer goods (for example, digital cameras), office supplies, computer ancillaries and cables, even land title documents, success rates are seldom above 50%.
The problem with 50% is that it isn’t adequate: there is so much manual intervention required to do the other half of the matching that it is more cost effective to do the whole thing by hand. While some manual intervention can be tolerated it can’t be at this level and you need a solution such as Silver Creek Systems’, where that degree of manual work is at a more reasonable level though, having said that, I do not know of any other vendor apart from Silver Creek Systems that can offer this sort of capability.
Silver Creek Systems’ product is called DataLens and it includes facilities to support semantically-based content profiling (that is, classifying records into content groups), standardisation (enforcing standards and normalising content), attribute identification, classification (aggregating data into taxonomies and schemas) and internationalisation (so you can have Spanish, Russian and other versions of the catalogue). The software can also operate in either real-time or batch mode, as appropriate.
I am not going to mince my words about DataLens: if you have a complex matching problem that goes beyond conventional name and address matching (not necessarily for products) then you must talk to Silver Creek Systems. To misquote a well-known beer advertisement: DataLens can get to parts of your matching problem that other data quality solutions cannot reach.