« Spiral 5–Prong Alligator Clip Memo Holder | Home | Clip–On Extendable Back Scratcher »

September 7, 2005

Automatic Sentiment Analysis — Software that analyzes the news and grades its tone

8ujs_1

Earlier this year Corpora Software of Surrey, U.K. released Sentiment, an analytical application for gauging the tone of news stories and then sorting them into one of three classes, each accompanied by a handy graphic indicator:

• A green smiley for positive

• A yellow blah face for neutral

• A red frowny face for negative

Stupid?

Perhaps — but consider that, as the company notes, human readers aren't perfect either, and human analysts are only able to process about 10 articles an hour while Sentiment can go through 10 every second — 36,000 an hour.

And Sentiment doesn't need coffee or bathroom breaks.

The company acknowledges that its machine isn't perfect but then, neither are humans.

For example, three expert readers are likely to agree about an article 85% of the time; about 90% of non–experts will agree with this consensus.

Sentiment agrees with the same expert consensus about 80% of the time.

Hjku_1

Here's a link to an article in New Scientist about this novel program.

September 7, 2005 at 02:01 PM | Permalink


TrackBack

TrackBack URL for this entry:
http://www.typepad.com/services/trackback/6a00d8341c5dea53ef00d83455be5d53ef

Listed below are links to weblogs that reference Automatic Sentiment Analysis — Software that analyzes the news and grades its tone:

Comments

Note really that hard.

Eudora use to have a 'chili filter' for their mail software...it used rudimentary Bayesian filters to detect the flame level of the message so as to warn you before ya sent something out that might potentially get you fired or lose a good friend or otherwise. Not hard at all...simply have to train the software as to what is good, bad, indifferent and after a while you take off the training wheels and let it go.

As for only an 80% agreeance level, that seems pretty low considering humans were getting 85 to 90. I found in my research, a well trained computer trained against a standardized rubric generally did better in the Rater::Computer than the Rater::Rater comparisons (then again, we were using 6 point scales as opposed to a 3 point -- and it was rating content as opposed to emotion).

Cool article though! Glad to see applications like this coming to the common man and not just locked away behind university firewalls!

Posted by: clif | Sep 7, 2005 5:24:49 PM

The comments to this entry are closed.