I’ve been researching sentiment analysis, and I think I’ve found pieces to suit a range of tastes and interests.
First the notes from a 2008 talk given by Lillan Lee from Cornell University. Lee’s topic is “…the ﬂood of interest in: sentiment analysis, opinion mining, and the computational treatment of subjective language.”
This is a good ‘who, what, why, how’, featuring:
Meanwhile, the epicly-named Dirk Shaw has a blog post from July 2009 asking Sentiment analysis, How much is good enough?. It’s a short post explaining the basic differences between manual and automated sentiment analysis.
In a post from June 2009, Irfan Kamal describes the approaches Ogilvy PR are taking to make sentiment analysis work for clients.
The Ogilvy post is useful in one respect because it shows the debate is still fledgling and the possibilities are wide open.
In another respect it helps illustrate that the people-or-software-or-both question of classifying sentiment is only one dimension of the problem. Sentiment analysis has to actually be of value to the people using the results.
In my view, the most useful output would be actionable insight, i.e. information that is directly useful for making decisions. That comes with a health-warning though: if we’re to make decisions on the basis of data, we had better be sure the data is completely valid, and leads to accurate conclusions.
It’s not clear yet that sentiment analysis will be able to deliver in such a concrete way, and at the level, for example, of government policy, sentiment data will find a place, but should be treated with healthy caution.
Marta Strickland has a post on these issues from September 2009. Focussing on product reviews, she identifies Five Reasons Sentiment Analysis Won’t Ever Be Enough, and concludes “What are we really trying to decide with this data? And are we asking the right questions?”
So it it hopeless then? Is sentiment analysis a turkey? Absolutely not. Asking critical questions absolutely indicates that smart people are taking sentiment analysis seriously, and figuring out how best to do it and how best to use it.
Rubber Republic has a working model for how we want to use sentiment analysis – for both our advertising and policy-making work. We’re evolving the model on the basis of trials, and we’re asking ourselves hard questions about how sentiment analysis can be valuable for our clients.
Enough about that (but more in future). Meanwhile, I’ll wrap up with a piece from each of the manual and automated perspectives. Nathan Gilliat has a September 2009 post on Scaling Human Analysis, while an April 2008 paper from Google discusses software approaches for Building a Sentiment Summarizer. Enjoy.
First posted at Team Rubber – Rubber Republic’s parent company