Taking Disambiguation Seriously

Today's search engines do a lot of things well, but one thing they don't do well - as astutely noted by erudite literary critic and National Post columnist Robert Fulford - is understand the precise meaning of the words you type. (Fulford's a crafty one. He was using Dogpile as early as 1997, and like any other sensible person wanting specific answers about search, he visits Search Engine Watch.)

I've been listening to the "disambiguation technology" story since 1999 when fledgling semantic technology companies like Oingo (now Applied Semantics) and Simpli (acquired by Netzero and then sold to pay-per-click keyword advertising service Search123) were launched. The examples are easy to grasp. A user types in "genesis," and it could mean dozens of things including a rock band, a biblical book, or a product. Java can mean coffee or a programming language.

I'm no linguist, but it's clear enough that there is a part of our brain where we file all these different meanings, storing them as, for all intents and purposes, different words. Insofar as a search engine is too stupid to help me disambiguate my query about "storage," for example, (I am searching for shelves, but a lot of the search results give me info about esoteric RAID computer disk storage), they do a poor job of putting the right kind of information in front of the user. They might even do a poor job putting the right ads in front of users, which could ultimately cost them money and frustrate advertisers. It makes little sense, for example, to subject sellers of low-margin shelving products to the higher prices that may be paid by advertisers wanting to advertise fancy RAID memory storage. But there is no way around it as long as the search engine does only "dumb" keyword matching. As far as our brains are concerned it's relatively inconsequential that the word storage is used to describe a shelf, and the same sequence of letters describes a system for filing terabytes of data. We don't see them as the same thing, although they might be related. What would be needed, in the ideal world, would be a unique code (or "word") for every substantially different concept and sub-concept. So the storage relating to data might be called "storbage" or something.

Well that might be impractical, but scientists *can* lend support to efforts to link researchers more closely with the concepts they're seeking. The old-fashioned way, the way that information science has done it for centuries, is to hard-wire an ontology (or classification system) and make everyone conform to it. Another approach, the more contemporary one, is to work with conceptual "meaning maps" to help classify information.

To be sure, end users can disambiguate their own queries to a certain extent. But should they be required to do all the heavy lifting of weeding out all the wrong meanings of words like java, storage, genesis, etc.? And will such efforts sometimes exclude pages because the phrases typed aren't matched on otherwise high-quality pages about the topic?

Given how diligently I've been trying to puzzle through all these issues, then, perhaps it isn't unfair or one-sided of me to reprint excerpts from a recent commentary by Applied Semantics - applied in this case as a critique of Google's much-ballyhooed Content Targeted Advertising.

"Google enters the contextual targeting advertising arena... Yahoo! upgrades its search technologies... Overture claims to have similar technology... So, what does all this mean?

"It means that there is an increasing demand and need for an innovative, durable contextual targeting solution for online advertisers. Major players like Google, Yahoo! and Overture are raising the stakes by entering the space. Online publishers and advertisers have to realize that a company's popularity, size and name recognition is not always indicative of its solution. In fact, many of these mega-company's contextual targeting applications are not effectively and efficiently getting the job done.

"Amongst these huge players is a small giant, packing the punch with the best contextual targeting product in the marketplace - AdSense. AdSense extracts the meaning of a web page and dynamically generates ads comprised of P4P (pay-for-placement) search terms and results on the fly. What makes AdSense technology different from the competition is its filtering capabilities, technology approach and partnership structure.

"In a head-to-head comparison between AdSense and Google's content targeting product, we found that AdSense has superior technology and obvious advantages in many categories:

• "AdSense technology processes actual content of web page, whereas Google technology is only based on URL/web log statistics
  
• "AdSense uses its CIRCA(tm) ontology to understand and extract key themes of a page, whereas Google relies solely upon user trends and patterns which are often inconsistent
  
• "AdSense aims to balance relevancy with CPC value to maximize effective CPM; Google has less effective CPM maximization
  
• "Ability to discern ambiguous terms - AdSense disambiguates, unlike Google who has no disambiguation capability
  
• "AdSense returns granular, precise keyword results; Google offers inconsistent results based on user search data; performs best on broad/general topics
  
• "AdSense provides advanced filtering technology; Google has no objectionable or competitive filtering mechanisms
  
• "AdSense offers customized ad designs tailored to customer preferences; Google has no ad design customization capabilities
  
• "AdSense offers a revenue share model, ensuring that we all share in the upside and are committed to developing the optimal implementation to ensure the highest user satisfaction."
  
  If Google's advertising program lacks semantic technology, then this hole must exist in its search technology as well.
  
  Whether it's brought about by a new era of robust metadata protocols, or technology such as Applied Semantics' CIRCA, it will be a pleasure to be able to type in the phrase "green jacket" and only receive results relating to the Masters golf tournament, or alternatively, only results NOT related to golf but rather to green jackets in general, according to my preference.
  
  And advertisers will be pleased the day that they can reach customers who are really interested in disk storage, for example, rather than having their ad show up on all kinds of pages relating to storage cabinets, just because some keywords happen to match.
  
  Disclaimer: Applied Semantics' comments about Google are mostly speculative.

  Posted by Andrew
  | | Permalink
  

   

  Canadian Golf Hero Trails Legendary Psycho in Buzz Index
  
  You can sure have a lot of pointless fun checking in with search trend trackers like the Yahoo! Buzz Index. Today, the #2 "mover," surging ahead over 500%, is Masters champ Mike Weir, narrowly beating out the #3 contender, Happy Gilmore (no guff). This is inexplicable. Much as I love Mike (another Canadian lefty golfer who signs autographs with his right hand, and grew up in Sarnia, where my sister was born), I like Adam Sandler's Gilmore character better. Who else would drop his imaginary gloves and cold-cock the overtanned Bob Barker while waiting to tee off? Almost makes up for the 12 Adam Sandler stinkers we've had to sit through.
  
  Even more inexplicable, though, is the fact that Weir was surpassed by Charles Manson. Don't tell me why. I don't want to know.

  Posted by Andrew
  | | Permalink
  

   

  Monday, April 14, 2003

  Everything You Wanted to Know about Black Holes
  
  On launching the latest phase of its Google Adwords advertising service, Google allowed that its new "content targeted advertising" might generate "slightly lower" clickthrough rates than search keyword advertising. Slightly? Holy guacamole, that might just win the next Search Engine Watch Award for Biggest Understatement of 2003.
  
  In more than a few cases, the CTR's for content targeted ads are so abysmally low as to be scarcely measurable. I just peeked in on a friend's long-running ad group that had a modest, but respectable, CTR of 0.8%, and which therefore generated a healthy number of daily clicks (around 60 per day at about 0.18 per click). The content targeted ads for the same group had a CTR of 0.0125%. That is, one in 8,000 viewers of the ad clicked on it. In all, three people clicked on the content-targeted ad today, generating a total of 0.15 in revenue for Google. Whoa, back up the dumptruck.
  
  Did Google fail to take adequate account of the very different dynamic of content-targeted advertising when they ran a limited pilot program last fall? Could it be... could it be!?!!?... that small text-based ads work well when people are in search mode, but very poorly in skyscraper creative next to content, to the extent that those AWFUL, INTRUSIVE, BLINKING ANNOYING GIANT CREATIVES we love to hate are actually more effective than Google's inobtrusive little text ads (the ones my clients know and love because search engine users actually click on them about 1% of the time, on average)?
  
  Hey, Google, these things are showing up on someone else's web site, not yours. Why not make the ads flash and stuff? Or generate some nice popunders? What's it to you if they lose a few readers? Just kidding, I know you'd never do this. Next thing they know people would be calling you a "media company." Bad enough that you already stand accused of being a portal.
  
  I'd expect Google to make quiet, restrained efforts to paper over these embarrassing stats by expanding the content targeting program to places where higher CTR's may be expected, such as, perhaps, email newsletters. Time will tell. If worst comes to worst, they can indoctrinate all of the users of the free version of Blogger into a new "product placement" scheme where they work Adwords ads "naturally" into their posts.

  Posted by Andrew
  | | Permalink
  

   

  Sunday, April 13, 2003

  Opera "Got Better," But Does it Matter?
  
  There are lots of new features in the latest version, 7.10, of Norway's pride and joy, Opera. But, it's much ado about nothing for me.
  
  I've tried it, but don't care much for it, and because not many browser extensions are supported, there is little incentive to use it. If you're looking for a better browser, go for Netcaptor. It does cost 30 bucks, but its handy features are worth every penny. Also, it's based on the IE engine, so you get all the benefits of IE with tons of additional features that you'll actually use! Lifetime upgrades and free, personal tech support are included.

  Posted by Cory
  | | Permalink
  

   

  Happy Tenth to the Browser
  
  An interesting article from the print edition of Newsweek ran on MSNBC.com about the 10-year anniversary of Mosaic, the first graphical web browser. Why was the web browser so important? Well, if you ever tried surfing the Internet with Gopher or any other text browsers, you would know instantly.
  
  Ah, I remember first using Gopher at the computer labs of the University of Missouri while a wee English student. Then, the labs deployed Mosaic, and I could suddenly surf the Web in vivid e-Technicolor. It was one small "click" for man, one giant "hit" for mankind.

  Posted by Cory
  | | Permalink
  

   

  Zeitgeist, Lies, and Videotape
  
  For a certain segment of marketers, the Lycos 50 and the like provide interesting info on hot trends.
  
  But for the vast majority of businesspeople, such as those attending my recent Nielsen Norman seminar, when you tell them that they can find out about the top 50 most popular keywords that week, they just shake their heads and say "what good is that"?
  
  So then I painted them a scenario - what if a major search engine would sell you detailed information about search patterns on as many topics as you want info on - say for $5,000 for a really detailed, real time report showing spikes in interest? They just echoed back to me what I'd already taught them in the morning session - why not just spend $100 piloting a Google Adwords or Overture campaign (or both), and get all the search frequency data you need for any term in your industry that you want to test?
  
  Never before has so much real time info about consumer behavior been available to so many businesses of all shapes and sizes. It's only a matter of time before people start gathering and selling this kind of data, or otherwise trying to capitalize on it. Oh wait, they already are: I received spam today that made its case for the usual member-expanding potions by trotting out the fact that "according to a major search engine, 15,000 people searched for the term penis enlargement last month, and 5,800 searched for the term sexual performance. Don't kid yourself bucko, she is more obsessed about it than you think!"... or something to that effect.
  
  Sudden bulges in consumer attention on hot topics might be interesting to track for mass marketers, or media looking for hot stories. These things have a lot of uses, obviously, but for those that can't use them, they're pretty much a distraction from the real business of reaching targeted customers.

  Posted by Andrew
  | | Permalink
  

   

  View Archived Posts