I don’t write off Wolfram Alpha, but I think the bar is now much higher. If Google’s results are close, then an incremental improvement won’t cut it. You can’t compete with good enough by just offering slightly better.

During the demonstration, Stephen Wolfram noted that 150 people are currently working on data input and tagging, with one of the surprises being the significant errors being found in published data as it’s cross-checked against other sources. He also mentioned that these human editors assist in creating rules for combining data sources, and in anticipating how to initially respond to ambiguous queries.

I think the trick to an effective next-generation query interface is 4-fold:

0. Support standard boolean, simple n-gram & new structured semantics within 1 query box with intuitive type-ahead. Additionally don’t be overly prescriptive as far as structure. I’ve only introduced 1 new piece of structure (single-quotes) to denote subject facets.

1. Align the structured query language with the facet map used for post search navigation.

2. Allow for “loose semantics” in the query predicates. So from my original example “mentioned” could be “talked about”.

3. Don’t be to smart for your own good, when in doubt fall-back to baseline queries. I believe to much data will be more welcome that result sets that don’t align well with the query.

In my case I have a bonus 5th factor which is the belief that regardless the query type (simple – structured) the results should also return an exploratory navigation structure which has the same facet topics as the structured query language. Full alignment between interface elements.

I started the first 2 years working diligently on a fully NLP driven query box and not only is it difficult to execute, users just don’t “get it”.

I’m much more encouraged with the testing of my structured query box with “simulated” (aka forgiving) natural language predicates. I’ll be very interested to see how it is received by you and other once it’s out there.

Thoughts?

1. At this time a completely open-ended NLP input system is neither practical or really needed. Couple that with the Google induced pavlovian response of how a query box should operate users don’t actually know how to use an NLP box to it’s fullest. I can attest to this via testing my query box with some very smart people and they had impedance issues.

Now even with all the negatives I think an NLP Query system that uses a controlled vocabulary that is aligned with a systems dB, entity or facet classes can provide an excellent jumping off point for getting a selected set of results back which can be further explored using the UI. For example let’s query the twitter stream with: “All ‘people’ who mentioned Wolfram Alpha ‘Today’”. In this example the words in ” are entity/facet/dB classes and the rest are the filters (which must take into account synonyms, etc. And again other facets must be returned so users can navigate the data further. What I’m suggesting here is a simple NLP query system that has a bit more smarts than good old boolean.

Even this is difficult to pull of effectively. I will have concrete test data to share as this is how I’m handling my query box, users can do the above or just use a basic query commands.

So in a nutshell my personal take is an integrated NLP Query box has value but the value comes not from supporting open ended questions but structured NLP queries. So not so much for general search but for data navigation.

Thoughts?