Sentiment Mining

There is no doubt that Mood News is a cool idea. I've been tracking it for a while as a member of the backstage community (from the beeb) - it was recently picked up by Infosthetics. However, it also illustrates the difficulty of developing this type of feature. The current most negatively ranked news is:

Algerian bomb suspect on trial

An Algerian man accused of involvement with bombings in Paris in 1995 goes on trial in France.

I would think that this was actually positive news. However, as the system uses keyword spotting to determine what is good and what is bad (bomb==bad), the orientation can escape the system. In addition, there is an issue of point of view. To whom is it good/bad news? Take a look at the current highly ranked positive news:

Palestinians hail EU aid deal

An EU aid package worth 120m euros is hailed by Palestinians and the US but condemned by Israel.

Sounds like bad news for Israel at least.

In analysing social media in the context of marketing intelligence, I tend to think of a number of categories of textual features, including:

• Sentiment: evaluative statements  (I like it, It sucked)
• Objective Favourable/Unfavourable statements (It is broken)
• Textual features: these are signals in the text that indicate some sort of mental state, especially emotion/affect (I can't wait to see it - indicating anticipation/excitement)

Generally, what we need to deliver is some mixture of these. The question raised by mood news is: what is the mixture appropriate for journalism? and, of course, what is of value to the user?

There is another set of considerations from the research side which needs to connect:

• Analysis that is useful to a customer/user of the output
• Psychological models of emotion/affect, etc.
• What is achievable with text analytics

That is a story for another post.

Comments

[Niall Cook]

Matthew, you hit on the single biggest problem in automated sentiment monitoring - the fact that sentiment is relative to the views of the reader. What is good news for one person/company/country may be bad news for another. So how can an automated system ever get it right?

[Matthew Hurst]

Niall - I'm not sure this is the biggest problem. I suspect that the real issue is how you interpret and report the results. If I read 'I liked Super Size Me', the author expresses the appreciation of the movie. To McDonalds, the company being criticised, this is not a good thing. However, by reporting the number of people who like the movie (positive sentiment), they can judge the *negative* impact to their brand.

The other aspect of mining textual data for emotional content (affect analysis) aims to capture the emotional state of the author - this is actually easier to do independent of the pov problem.