Everybody loves a good snog – Finding sentiment in the UK via Twitter | Hacker News Reader

Everybody loves a good snog – Finding sentiment in the UK via Twitter

rawkes.com

16 pointsrobhawkes9 comments

As a commenter on the article has already asked, it'd be nice to have an explanation on how the sentiment is actually determined.

I'll be writing this up in more detail soon, but the sentiment is basically derived through the ANEW dataset. This dataset scores a few thousand English words with sentiment values from 1 (very unhappy) to 9 (very happy). I find the sentiment value for each word in a tweet and then average that value by the amount of words in a tweet to get the sentiment value for each tweet.

This seems a rather simplistic approach when analysing sentiment in a country known the world over for truly impenetrable sarcasm, and the application of irony so blunt that it often boils down to simply stating the exact opposite to what you actually mean, often with a total absence of syntactic cues that this is happening.

The humour that brits derive from this often seems like it is defined by fact that these cues aren't there at all. In short, I think you'd have to do a lot of quite complex higher order link analysis to determine people's true opinions.

In this example, the royal wedding, imagine the following tweet: "This is a fantastic and important day for britain, a fabulous use of public money, I feel so proud to celebrate such a deserving couple!!"

Put it like this, I wouldn't bet money on a simple word scoring algorithm getting the intended sentiment correct here.

Thanks for your input Jonnie, you've highlighted one of the main limitations of text-based sentiment analysis. I've never always asserted that the sentiment derived from this study is not indicitive of "true" sentiment of an individual person. Rather, the sentiment derived here is simply saying that people on Twitter are using particular words that connotate happiness or sadness.

This is the same limitation that most other studies have found, but it doesn't make the results any less interesting.

Another way to look at it is as signal-to-noise. Whatever stereotypes of the UK exists, we don't all talk sarcastically (at least not all the time). Because of this the majority of tweets, which are probably not sarcastic, average out the hard to read tweets. Again, this is something that other studies have found as well.

My aim here isn't to fault the method of sentiment analysis. I'll leave that to the guys at Florida University who created it. :)

Oh I didn't mean to criticise anything you're doing. There's no such thing as bad data mining, only bad analysis of the results :)

I just worry that people in our governments and security services will sieze on work like this which is largely being carried out for the amusement and satisfaction of academics and general geeks, and decide that it is appropriate to make specific judgements about specific people based on the data.

It doesn't help that I read this article just after I woke up this morning:

http://www.guardian.co.uk/uk/2011/may/03/protester-sue-polic...

Combine it with this recent episode: (the guy was eventually convicted and fined)

http://www.guardian.co.uk/world/2010/jan/18/robin-hood-airpo...

...and with all the recent hype over 'cyberwar,' I start feeling a bit chilly.

I agree completely. These results shouldn't be simply taken at face value if they're to be used for further analysis.

Really interesting Rob! Well done.

This information would be much more valuble with data for a "normal" day. For example, Is the main bump around noon just a regular lunch high?

You're right. I haven't been able to run every type of comparison yet, but in the meantime you could check out the other graphs that I've made that cover normal days: http://www.flickr.com/photos/robhawkes/sets/7215762573356079...

Reply on news.ycombinator.com