Ask HN: How to determine temperature of text?
4 comments
The usual approach for these problems is to admit that you don't know what the rules are, and instead train a statistical classifier based on some sample data:
http://crm114.sourceforge.net/
Depending on how hard your problem is, the accuracy of the classifier may not be very good. You can try to come up with fancier "features" to extract from the items and feed to the classifier, or you can use "artificial artifical intelligence" and just farm out the work to Mechanical Turk at a penny per review.
http://crm114.sourceforge.net/
Depending on how hard your problem is, the accuracy of the classifier may not be very good. You can try to come up with fancier "features" to extract from the items and feed to the classifier, or you can use "artificial artifical intelligence" and just farm out the work to Mechanical Turk at a penny per review.
Well outside my field, but I believe this is a reasonably well-established problem in text processing/NLP, known as "Sentiment Analysis" -
http://en.wikipedia.org/wiki/Sentiment_analysis
My very brief google scholar search on that key phrase turned up a whole lot of results that look like they may provide fairly concrete algorithms and techniques for doing what you describe.
http://en.wikipedia.org/wiki/Sentiment_analysis
My very brief google scholar search on that key phrase turned up a whole lot of results that look like they may provide fairly concrete algorithms and techniques for doing what you describe.
As other posters have said, it's a genuinely hard problem that touches on natural language quantification. You can find an excellent resource on the relationship between english words here http://wordnet.princeton.edu/. You could manually rank a set of adjectives and then use wordnet to find synonyms and generate some heuristic from there, but the bottom line is I think it's going to be a huge time sink unless you get someone with NLP experience involved.
Thank you all for your wonderful comments. I now have some concrete directions to look. I must say it is much harder than I initially thought.
Are there companies employing any of these techniques to provide some service? This seems to be the next step up from keyword-based search. I remember seeing some startups providing a gauge of positive press coverage on stock symbols. How do they do it?
Are there companies employing any of these techniques to provide some service? This seems to be the next step up from keyword-based search. I remember seeing some startups providing a gauge of positive press coverage on stock symbols. How do they do it?
For one example, I've seen a couple of sites that classify Twitter posts about movies as positive or negative. I don't know what techniques they're using.
http://flixpulse.com/
http://www.twittercritic.com/
http://flixpulse.com/
http://www.twittercritic.com/
Just checked out the 2 sites. They don't seem to be active? The movies there seem to be dated.
Actually flixpulse.com classified some reviews wrong. Here are some of the "bad" reviews on "Bolt":
-- "Bolt" is bringing AWESOME back -- @seanbonner awesome! BOLT loves you back.
Whatever algo they use need some tuning. But I don't think anyone is actively working on it anymore.
Actually flixpulse.com classified some reviews wrong. Here are some of the "bad" reviews on "Bolt":
-- "Bolt" is bringing AWESOME back -- @seanbonner awesome! BOLT loves you back.
Whatever algo they use need some tuning. But I don't think anyone is actively working on it anymore.
The harder question: now that we know the review is positive, how can we assign a degree of positiveness? "The best since sliced bread" is much more positive than "good, but can be better".