I’ve gotten a ton of response to yesterday’s piece on the hidden bias of sentiment analysis (thank you!) with a lot of folks chiming in one way or the other on the current state of automated sentiment analysis. Some of you, quite correctly, sent me messages that went something like: “OK, smarty-pants, how would YOU do it?” I can’t give you the smarty-pants answer to this, but I CAN give you one alternative to tackle the problem today.
What I am about to describe, it should be noted, does NOT apply to other forms of text analysis–if you want to analyze a massive body of brand mentions for content, or to identify clusters or segments by subject, or customer service opportunities, etc., there are loads of great tools for this, and they all work fine. I’ve used everything from SPSS Text Analytics tools on large datasets, to basic concordance software and a copy of Tinderbox on small sets, and they all do what they say on the tin. There is a great need to automate that kind of work, and great tools to get that particular job done.
No, I’m talking specifically here about sentiment analysis–not what the brand mentions were about, but how the “mentioners” felt about the brand. There is an apocryphal story about the development of the Fisher Space Pen, which was developed literally so astronauts could write in space without their pens exploding or the ink not flowing. Hundreds of thousands of dollars were spent licking this problem, which led to the development of the Fisher Space Pen, a pressurized, sealed pen capable of writing in the most extreme environments. Meanwhile, so the story goes, the Russians licked the same problem by sending their cosmonauts up with pencils.
My method will appeal to those of you who appreciate the latter solution.
Step One: Pick a timeframe, and add up all of the brand mentions you have in your dataset for that timeframe.
Step Two: If there are less than a thousand, brew a pot of coffee, set aside a couple of hours, and sharpen your pencil (I am a big fan of the pencil.) Do it by hand. Find some friends, and it will take less than an hour. You’ll be nearly 100% accurate, it won’t take as long as you think, and you’ll know the data cold. It will feel good. Trust me.
If there are more than a thousand brand mentions (which is likely for a Fortune 1000 company), then you need an intermediary step between the two just mentioned:
Step One-And-A-Half: Take a random sample of 1,000 brand mentions across a proportionate mix of social media channels.
Then, you can proceed to Step Two, brewing the coffee and sharpening the pencil, etc.
See, if you have a few hundred or even a couple of thousand mentions, then by all means, take a census. But if you have more than that–the only census that matters in this country is done every ten years and will cost over 15 billion dollars in 2010. You don’t need a census–lucky you–because sampling works. Sampling is the law (I like to say that in a Sly Stallone voice). If I take 1,000 randomly selected samples from a population of 10,000 mentions, then I could repeat the exercise 100 times, and 95 of those would put my sentiment measure within about 3% either way. Increase the total population to a million, and it’s still about +/- 3%, 95% of the time. It’s reassuringly like magic.
Is it perfect? No. First of all, while it’s easy to generate a random sample within a given social media channel (tweets, for instance), it’s a bit tricker to do the same across all social media channels and ensure that the sample is not only random but representative. Also, five percent of the time (the confidence interval I selected to get you +/- 3% at n=1,000) the data will not be within three percent either way. It could be really, disastrously wrong. But so could the alternatives.
Look, I’m no Luddite. Automated sentiment analysis is getting better–lots of folks trying to crack that nut have left passionate comments on this blog, and I have no doubt that they are getting closer and closer. Some of them are getting very close indeed. There is no perfect solution–I’m merely presenting an alternative.
With that, I throw open the door to you–how does your company measure sentiment today? Are you? What kind of success have you had? And–most importantly–how have you used that data? I’d love to hear your stories.