Posted: 03 Aug 2013 07:20 AM PDT
I read “How Can I Pass the String ‘Null’ through WSDL (SOAP)…” My hunch is that only a handful of folks will dig into this issue. Most senior managers buy the baloney generated by search and content processing. Yesterday I reviewed for one of the outfits publishing my “real” (for fee) columns a slide deck stuff full of “all’s” and “every’s”. The message was that this particular modern system which boasted a hefty price tag could do just about anything one wanted with flows of content.
Happily overlooked was the problem of a person with a wonky name. Case in point: “Null”. The link from Hacker News to the Stackoverflow item gathered a couple of hundred comments. You can find these here. If you are involved in one of the next-generation, super-wonderful content processing systems, you may find a few minutes with the comments interesting and possibly helpful.
My scan of the comments plus the code in the “How Can I” post underscored the disconnect between what people believe a system can do and what a here-and-now system can actually do. Marketers say one thing, buyers believe another, and the installed software does something completely different.
- A person’s name—in this case ‘Null’—cannot be located in a search system. With all the hoo-hah about Fancy Dan systems, is this issue with a named entity important? I think it is because it means that certain entities may not be findable without expensive, time-consuming human curation and indexing. Oh, oh.
- Non English names pose additional problems. Migrating a name in one language into a string that a native speaker of a different language can understand introduces some problems. Instead of finding one person, the system finds multiple people. Looking for a batch of 50 people each incorrectly identified during processing generates a lot of names which guarantees more work for expensive humans or many, many false drops. Operate this type of entity extraction system a number of times and one generates so much work there is not enough money or people to figure out what’s what. Oh, oh.
- Validating named entities requires considerable work. Knowledgebases today are “built automatically and on-the-fly. Rules are no longer created by humans. Rules, like some of Google’s “janitor” technology, figure out the rules themselves and then “workers” modify those rules on-the-fly. So what happens when errors are introduced via “rules.” The system keeps on truckin’. Anyone who has worked through fixing up the known tags from an smart system like Autonomy IDOL knows that degradation can set in when the training set does not represent the actual content flow. Any wonder why precision and recall scores have not improved too much in the last 20 years? Oh, oh.
I think this item about “Null” highlights the very real and important problems with assumptions about automated content processing. Whether the corpus is a telephone directory with a handful of names or the mind-boggling flows which stream from various content channels.
Buying does not solve long-standing, complicated problems in text processing. Fast talk like that which appears in some of the Search Wizards Speak interviews does not change the false drop problem.
So what does this mean for vendors of Fancy Dan systems? Ignorance on the part of buyers is one reason why deals may close. What does this mean for users of systems which generate false drops and dependent reports which are off base? Ignorance on the part of users makes it easy to use “good enough” information to make important decisions.
Stephen E Arnold, August 3, 2013
Sponsored by Xenky