Powerset, a start-up search company focused on so-called “natural language search” is now open for business (see Techcrunch story).    So far, you can search only the content on Wikipedia.  There’s a reason for that: the Powerset technology doesn’t simply index content by keyword; rather, it runs more computationally intensive software that attempts to figure out the semantic meaning of the content, and then make that content searchable with a natural language interface.   So, despite $12.5M of VC funding, they don’t have the computational resources to index the web.   Sounds great to you?  Well, I’m not so sure it is…

Unfortunately, I think Powerset has two major problems: firstly, the kind of approach Powerset appears to be taking performs worse than keyword search; and secondly, Google is, in a low-key way, already implementing a “natural language” approach to search that does work better than keyword search.

Back in the late 1990s, I developed a natural language search system, based on a model of how the human brain understands language, for use in the biotechnology and pharmaceutical industries (you can read a bit about it in this publication on the American Chemical Society web-site - it’s an easy read, with no overly technical content in it!).     The key thing that made our system work, was that we built in detailed knowledge about the particular domain we were interested in, and the system was designed from the ground up to make use of domain knowledge.    General purpose, or domain-neutral, approaches to natural language understanding software, by themselves, simply don’t work very well compared to keyword search.

What does this mean in practice? Well, when you type in your natural language query into Powerset, it means you’re unlikely to get a clear answer to your question, because Powerset’s domain-neutral software doesn’t really understand the question.   However, if you type in a natural language query into Google, and it’s specific type of question that the Google software understands, then you get a great answer back.    Here’s an example query -  “What is Elton John’s real name?”

Powerset Results

On Powerset, the results are as follows:

1. Elton John

   (A spoonerism for “Budget Rent-O-Van” and a reference to his real name Reg.) … Spots related to Elton John’s childhood and career

2. close

   Elton John discography

   Elton John’s Milestones (1980) … | 1994 | “Ain’t Nothing Like the Real Thing” (with Marcella Detroit) (Nickolas Ashford/Valerie Simpson) | Duets | 24 | (N/R) | (N/R) | – | – |

3. close

   Elton John’s Greatest Hits

   Though the album includes many hits from Elton John’s early era, it is noted for leaving off several of his more popular songs.

4. close

   List of albums (E)

   Elton John – Elton John’s Greatest Hits Volume II (1977) … Evermore – Real Life (2006)

5. close

   Elton John’s Greatest Hits Volume II

   Elton John chronology Blue Moves (1976) Elton John’s Greatest Hits, Vol. 2 (1977) A Single Man (1978)

Not exactly what someone asking that question is looking for.  It’s worse than a keyword search.

Google Results

On the other hand, if you type the same question into Google, you get the following answer:

Elton John — Birth Name: Reginald Kenneth Dwight

According to http://www.movietome.com/people/3666/elton-john/index.html - More sources »

That’s a perfect answer.  Google’s software understood what I was asking, and gave me a clear answer back.    How did it understand the question?  It didn’t take a domain-neutral “subject verb object” style approach to trying to understand the question.  Instead, the software “knows” that a search with “the name of a person” and the phrase “real name” is most likely to mean, “What is this person’s birth name?”.

I really think Powerset has a lot of work to do if it wants to be taken seriously as a player in search.   The truth is, as of now, Google gives better results for most searches, including some natural language searches.    That’s not to say Powerset can’t find a niche where it can succeed - but, in my opinion, promoting general natural language queries isn’t the right way forward for the company.