My Photo

« ReadBurner Burner | Main | Bush Versus The Election »

January 23, 2008

Comments

Peter Turney

With just slightly more complicated queries, you could get an estimate of whether the buzz is positive or negative:

http://arxiv.org/abs/cs.CL/0309034

Matthew Hurst

Peter,

I disagree. Determining sentiment requires generating tuples of the form (speaker, target, polarity). While determining candidate targets and polarity may be possible, the association is a hard problem. Many approaches to this make a simple sentence assumption: a sentence mentioning Obama which is also positive is positive about Obama. Unfortunately, this has many problems. Sure, in some domains it works out well, but in the political domain - where there are many targets and often opposing opinions - things are harder.

I do believe that it is possible to create a system which does a good job of sentiment analysis for politics. But I don't believe it can be done with 'slightly more complicated queries.'

That being said - I'd be happy to be proven wrong. Could you show some examples of the queries you have in mind and their accuracy?

Peter Turney

"Many approaches to this make a simple sentence assumption: a sentence mentioning Obama which is also positive is positive about Obama."

I expect that this assumption works well enough, given large sample sizes. I assume the errors would cancel out, if you take the average of a big sample. But you're right, it is a big assumption, and I really don't know how well it would work.

Kevin

Now how about those graphs without Ron Paul? That's what I'd like to see.

Matthew Hurst

Removing the term 'paul' appears to have a pretty significant impact.

http://tinyurl.com/yrfx9v

The comments to this entry are closed.

Twitter Updates

    follow me on Twitter

    March 2016

    Sun Mon Tue Wed Thu Fri Sat
        1 2 3 4 5
    6 7 8 9 10 11 12
    13 14 15 16 17 18 19
    20 21 22 23 24 25 26
    27 28 29 30 31    

    Categories

    Blog powered by Typepad