Blank (Media) Slate

I was finally accepted into the PowerSet PowerLabs beta program to give their natural language search technology a whirl. As far as I can tell, they’re not in “beta” or even “alpha” for any specific product line right now. Instead, they’ve released some demoware showcasing some of their capabilities.

Among the demos they’ve put up are structured queries for content related to business, arts, and quotes. Each of these allows the user to select one of a dozen or so canned query formats into which free-form nouns and/or verbs can be placed. The input for each of these canned queries is undoubtedly wrapped in some magic prior to execution against the specified source. By restricting user input, and directing the query toward engines tuned for the specific need they can narrow down the degrees of freedom.

For example, one of the business searches you can run is in the form “Who acquired INSERT:COMPANY”. Entering a company name will bring up a list of results highlighting the assumed relationships as found in articles on Wikipedia (their only current source for information).

Their goal is to home in on the intent of the user, and provide results that are better than standard keyword searching. To allow users to judge the quality of the results, most of the demos reply with side-by-side comparisons between what PowerSet can do next to what is returned by the inputs as keywords.

The demo that provides the most flexible input from the user is their PowerMouse application. Using it the user is able to build an undirected (ie. not forced to their “business, arts, or quotes” categories) query in the format “subject-verb-subject”. In fact, you can leave one (or two) of the fields blank to see what it finds. One of their canned examples is “zombies - eat - BLANK”. It is gratifying, then, to see Wikipedia articles returned that include “zombies - eat - brain” (along with eating “body part, boy, chick, debbie, flashback, franklin, galactus, granddaughter, hawkeye, head, man, meat, member, neighbor, people , richards, schoolchild, shell, study, sullivan, team, vet, yoshi”).

To try it out myself, I wanted to see what it would pick up about my friend Sunita Williams running the Boston Marathon while aboard the International Space Station. I already knew that there was a note about this on her Wikipedia bio page, so I figured it’d be a slow ball for PowerSet to knock out of the park.

I entered the query: “Sunita - ran - BLANK”

The result set included the following:

1. Sunita - dump - maya
2. Sunita - tie - maya
3. Sunita - survive - ordeal
4. Sunita - seek - reside
5. Sunita - develop - feeling
6. Sunita - get - marry
7. Sunita - go - look
8. Sunita - set - world record

I’m not sure what synset graph they’re using for “ran” in this context, but the first seven results were clear misses. The last entry, though, made me think PowerMouse hit upon what I expected as it was able to find Sunita’s bio page (as opposed to the unrelated “Sunita Parekh”, a TV soap opera character). Unfortunately, however, expanding the results showed that the blurb it keyed off was actually about her record-breaking space walk (no mention of the marathon).

To make sure I was remembering her bio page correctly, I checked and here’s the paragraph mentioning the marathon:

On April 16, 2007, she ran the first marathon by an astronaut in orbit.[6] Williams finished the Boston Marathon in four hours and 24 minutes.[7][8] The other crew members reportedly cheered her on and gave her oranges during the race. Williams’ sister, Dina Pandya, and fellow astronaut Karen L. Nyberg ran the marathon on Earth, and Williams received updates on their progress from Mission Control.

I’m surprised it didn’t clue into the part of the sentence that reads “she ran the first marathon”. That seems to be about as clear a match for the query as could be expected in many reasonable situations.

It’s too bad that the author of the blurb led the paragraph with a pronoun rather than Sunita (or Williams). It’s possible that the pronoun recursion required to connect the noun was unable to detect the association. Possibly compounding the problem is that the nearest previous noun was “Joan Higginbotham”.

I wonder, then, if the query would have picked it up had the sentence read “Williams ran the first marathon” (or more specifically, “Sunita Williams ran the first marathon”). Since it’d be better encyclopedia formatting to lead the paragraph with her last name, and updating the page wouldn’t hurt, I have half a mind to edit it and try the query again.

The shining star in the experience, though, was the user interface of their support site. Very nice use of in-situ form editing and feature flow. Not great knowledge repository, but fun to play with. My hope is they’re engaging with a good group of demo testers during their shakeout cruise. I look forward to watching as the training wheels come off to see how it works in the wild.