Saturday, January 7, 2006

Lexxe

Lexxe (currently in Alpha... a new trend?) claims to apply Natural Language Processing on questions you enter to come up with an answer. It performed quite well on “Who is the chancellor of Germany?” – a question to which Google’s Q&A feature wrongly replies Gerhard Schröder* – but it also failed me on many other questions. [Thanks Miel.]

*As of late 2005, it’s Angela Merkel.

By the way, here’s an approach to finding answers, as previously discussed (I removed this feature from FindForward not long after Google’s Q&A was released):

1. Given a question, remove words like “how”, “what”, “why”,
2. Use Google to find the most popular arrangement of words contained in the sentence by going through all permutations,
3. Use the result as pattern to match on the web,
4. Output a nicely cut snippet.

This algorithm is incredibly simple, but step 2 can be time-consuming to process depending on the search API used. Here’s an example:

1. The user enters “Who is the chancellor of Germany?”
2. This is permutated into “the is germany chancellor of”, “germany is the of chancellor” and so on.
3. Finally – theoretically – “the chancellor of germany is” returns the most results (at the moment, it has a page count of 156 in Google). We disregard the original phrase in our search.
4. We query Google for “The chancellor of Germany is * * *” and copy the bold words to get our answer. (Well, in most cases... this works best for those kind of questions which are answered often on the web.)

Lexxe by Philipp Lenssen | Comments (10)

>> More posts

Advertisement