TrueKnowledge, NLP, and HCI

It's exciting to see NLP present in so many applications these days. Some of the companies I follow and have a keen interest in seeing develop are: Attensity, Powerset, Hakia, and RadarNetworks. But out of the blue last week ( and I'm still not sure how I missed this as I'm usually on top of companies like this ) comes TrueKnowledge. If you haven't peeped their demo yet, you should do so. The site looks great but it's in "private beta" right now. I'm looking forward to testing their tech out once it's finally released.

TK bills themselves as a site for "Question-Answering" and "Enhanced Search". Based on the demo, you can type in a question in English such as, "Who invented Perceptrons?", and it will provide an Answer. It's not just using keywords, it does some parsing of the sentence you enter and then applies some inference rules. In the demo, they ask "What time is it at Google?", and the correct time is returned. Very cool stuff.

So I like asking a question and getting a straight answer, but I have to critique how the answer is returned. This doesn't apply to just TK either. I see other companies doing this, and it's because Engineers come up with this stuff. When we have the Engineer hat on, we do brilliant things, but forget about the little details like user experience. So...

In the demo, when you ask a question of their site, you get a straight answer.

Example: You ask "is J Lo single?", the site returns "No" and a link it thinks will back up it's answer. That parts awesome.

But then there's all this other verbose stuff with the answer. For the same question, the site also returns: "There are 2 defintions of "single" , then provides the 2 definitions,
then goes on to say "I took the meaning of your question to be.." and inserts something from the definitions previously listed. There can then be several links that represent the "inference path" for how the machine came up with the answer it's returning. In their Google-time example, a ton of links are returned.

Something about this nags at me that the HCI or UI design is off somewhere. It looks like they just jammed a bunch of System.outs in their code and blurt out the logging in the answer. They try to gussy it up and anthropomorphize it to make the logging seem helpful, but I'm not sure it really is. Maybe they should have a verbose mode, that will give you all that extra crap if you want it, and a trusted mode, where you just trust TK is giving you the right answer. Or maybe in trusted mode you get a straight answer and a confidence percentage. If the user questions the answer returned or the confidence rating, they can always go into verbose mode to validate the response, but I don't think logging should be on by default. Anyway, for me, I'll dig all that extra info, but I wonder if the average user will, or how all those inference links will look jammed on a cel phone.

Anyway. Back in 2001, I'd type a query in Google and it was like it could read my mind, I loved it. Their search just blew me away. Nowadays, there are a bunch of search engines out there, and I try to check them out occasionally, but I find myself going back to the Goog regularly cause "The devil I know....". But I'm almost always assured that the first few links will be trying to sell me items from ebay or amazon.

I'm all for a site that I can query that will return answers and not links. More power to 'em!