• Home
• About
• Writers
• Landscape
• Wizards Index
• Fraud
• Wanna Be Like Larry?

Semantic Search: The View from a Taxonomy Consultant

My team and I are working on a new project. With our Overflight system, we have an archive of memorable and not so memorable factoids about search and content processing. One of the goslings who was actually working yesterday asked me, “Do you recall this presentation?”

The presentation was “Implementing Semantic Search in the Enterprise,” created in 2009, which works out to six years ago. I did not recall the presentation. But the title evoked an image in my mind like this:

I asked, “How is this germane to our present project?’

The reply the gosling quacked was, “Semantic search means taxonomy.” The gosling enjoined me to examine this impressive looking diagram:

Okay.

I don’t want a document. I don’t want formatted content. I don’t want unformatted content. I want on point results I can use. To illustrate the gap between dumping a document on my lap and presenting some useful, look at this visualization from Geofeedia:

The idea is that a person can draw a shape on a map, see the real time content flowing via mobile devices, and look at a particular object. There are search tools and other utilities. The user of this Geofeedia technology examines information in a manner that does not produce a document to read. Sure, a user can read a tweet, but the focus is on understanding information, regardless of type, in a particular context in real time. There is a classification system operating in the plumbing of this system, but the key point is the functionality, not the fact that a consulting firm specializing in taxonomies is making a taxonomy the Alpha and the Omega of an information access system.

The deck starts with the premise that semantic search pivots on a taxonomy. The idea is that a “categorization scheme” makes it possible to index a document even though the words in the document may be the words in the taxonomy.

For me, the slide deck’s argument was off kilter. The mixing up of a term list and semantic search is the evidence of a Rube Goldberg approach to a quite important task: Accessing needed information in a useful, actionable way. Frankly, I think that dumping buzzwords into slide decks creates more confusion when focus and accuracy are essential.

At lunch the goslings and I flipped through the PowerPoint deck which is available via LinkedIn Slideshare. You may have to register to view the PowerPoint deck. I am never clear about what is viewable, what’s downloadable, and what’s on Slideshare. LinkedIn has its real estate, publishing, and personnel businesses to which to attend, so search and retrieval is obviously not a priority. The entire experience was superficially amusing but on a more profound level quite disturbing. No wonder enterprise search implementations careen in a swamp of cost overruns and angry users.

Now creating taxonomies or what I call controlled term lists can a darned exciting process. If one goes the human route, there are discussions about what term maps to what word or phrase. Think buzz group and discussion group and online collaboration. What terms go with what other terms. In the good old days, these term lists were crafted by subject matter and indexing specialists. For example, the guts of the ABI/INFORM classification coding terms originated in the 1981-1982 period and was the product of more than 14 individuals, one advisor (the now deceased Betty Eddison), and the begrudging assistance of the Courier Journal’s information technology department which performed analyses of the index terms and key words in the ABI/INFORM database. The classification system was reasonably, and it was licensed by the Royal Bank of Canada, IBM, and some other savvy outfits for their own indexing projects.

As you might know, investing two years in human and some machine inputs was an expensive proposition. It was the initial step in the reindexing of the ABI/INFORM database, which at the time was one of the go to sources of high value business and management information culled from more than 800 publications worldwide.

The only problem I have with the slide deck’s making a taxonomy a key concept is that one cannot craft a taxonomy without knowing what one is indexing. For example, you have a flow of content through and into an organization. In a business engaged in the manufacture of laboratory equipment, there will be a wide range of information. There will be unstructured information like Word documents prepared by wild eyed marketing associates. There will be legal documents artfully copied and pasted together from boiler plate. There will be images of the products themselves. There will be databases containing the names of customers, prospects, suppliers, and consultants. There will be information that employees download from the Internet or tote into the organization on a storage device.

The key concept of a taxonomy has to be anchored in reality, not an external term list like those which used to be provided by Oracle  for certain vertical markets. In short, the time and cost of processing these items of information so that confidentiality is not breached is likely to make the organization’s accountant sit up and take notice.

Today many vendors assert that their systems can intelligently, automatically, and rapidly develop a taxonomy for an organization. I suggest you read the fine print. Even the whizziest taxonomy generator is going to require some baby sitting. To get a sense of what is required, track down an experienced licensee of the Autonomy IDOL system. There is a training period which requires a cohesive corpus of representative source material. Sorry, no images or videos accepted but the existing image and video metadata can be processed. Once the system is trained, then it is run against a test set of content. The results are examined by a human who knows what he or she is doing, and then the system is tuned. After the smart system runs for a few days, the human inspects and calibrates. The idea is that as content flows through the system  and periodic tweaks are made, the system becomes smarter. In reality, indexing drift creeps in. In effect, the smart software never strays too far from the human subject matter experts riding herd on algorithms.

The problem exists even when there is a relatively stable core of technical terminology. The content of a lab gear manufacturer is many times greater than the problem of a company focusing on a specific branch of engineering, science, technology, or medicine. Indexing Halliburton nuclear energy information is trivial when compared to indexing more generalized business content like that found in ABI/INFORM or the typical services organization today.

I agree that a controlled term list is important. One cannot easily resolve entities unless there is a combination of automated processes and look up lists. An example is figuring out if a reference to I.B.M., Big Blue, or Armonk is a reference to the much loved marketers of Watson. Now handle a transliterated name like Anwar al-Awlaki and its variants. This type of indexing is quite important. Get it wrong and one cannot find information germane to a query. When one is investigating aliases used by bad actors, an error can become a bad day for some folks.

The remainder of the slide deck rides the taxonomy pony into the sunset. When one looks at the information created 72 months ago, it is easy for me to understand why enterprise search and content processing has become a “oh, my goodness” problem in many organizations. I think that a mid sized company would grind to a halt if it needed a controlled vocabulary which matched today’s content flows.

My take away from the slide deck is easy to summarize: The lesson is that putting the cart before the horse won’t get enterprise where it must go to retain credibility and deliver utility.

Stephen E Arnold, May 9, 2015

Comments

Copyright © 2008 Beyond Search · ArnoldIT.com · Disclaimer · Revolution News theme by Brian Gardner · Log in