Moodteller: Estimating mood levels in LiveJournal

Moodteller: Estimating mood levels in LiveJournal

What is this?

A demo of a technique for estimating the current mood levels in the blogosphere using the language used by bloggers.

LiveJournal enables its users to tag their posts with mood indicators. We currently analyze approximately 5000 LiveJournal blog posts per hour using statistical language processing methods and estimate, according to the textual features of the posts, the percentage of them which are "happy", "sad", "excited", and so on; we do not use the mood indication given by LiveJournal. After the estimation is done, we check how good it is by examining the real indications given by the bloggers. The graph above shows the estimated and actual values per mood.

What do the numbers mean?
The Y-axis shows the number of LiveJournal posts reported with the given mood per hour. Accuracy is defined as (estimated-value/actual-value), and the average is done over all hours appearing in the graph. The correleation is the statistical correleation coefficient between the estimated and actual graphs.

Lots of additional technical information about how this works can be obtained here (PDF file): Experiments with Mood Classification in Blog Posts, or by contacting the author.
Why do this?

Why do we need to estimate the mood levels of LiveJournal, if we can just track the actual levels?
- First, we use LiveJournal only to show the accuracy of our method, since with this data it can be measured; but the same technique can be applied to other blog posts, not tagged with moods -- or even non-blog domains.
- Second, by discovering a large gap between the estimated and actual values (a sudden drop in accuracy), we can "sense" that something is going on -- there is a change in the pattern of language used for a certain mood, typically meaning that a global event is being reflected in a large amount of blog posts. This happened, for example, following Hurricane Katrina.