Short answer is probably yes (if we are talking about the next election). But not as it was done in this election cycle. But there is a lot of potential.
Here is what we can answer: we can tell what people are talking about; categorize all comments into meaningful contextual buckets that we call Topics and do trend analysis based on that. We can also accurately extract names, places and brands.
If we could do this again the results would be dramatically more accurate and actionable
This is how we would do this next time: First we would (automatically) deduce commenters political affiliation from their first comment. At least in this type of situation of two candidates, that would be fairly easy. Whether people like this is a whole other question but we could do it anonymously by giving each commenter an ID. (That is how we handle the feedback analysis. We take away all the personal information.) Knowing commenters political affiliation is important: it enables us to fix the “cross commenting” problem so that we could really analyze the supporters and antagonists compared to their political affiliation. This would enable us to find out what each candidates supporters are talking about in the candidate’s and opposition’s site.
I wish Facebook had a location API
One really cool thing to have would be commenters location by state or preferably city. With this type of volumes, it would make the geographic visualization meaningful. You could (maybe) see clusters of certain topics form in certain areas. This would of course mean that Facebook would make the location available through their API, which we think will NEVER happen (because it is one of their biggest assets). One option could be to develop a Facebook Political Analysis app within the FB ecosystem. However, I am not sure whether they make the location available even then.
With minor optimization work sentiment analysis would be much more accurate
Etuma sentiment analysis is very accurate for those industries (and languages) that we have optimimized it for. Because we did not have any data nor the resources to do the optimization, we never reached the sentiment analysis accuracy we would have like to. Next time we would use this data to run a few man working week sentiment optimization project. We have a clear process and tools for it. The sentiment accuracy would be in totally different level. Right now we estimate that it was about 70% (what this mean is that in 7 Topic mentions out of ten we can accurately say whether people are talking about the Topic negatively, neutrally or positively). With optimization it would get close to 90%. Then you could really rely on the statistical trend analysis and detect even minute changes in Topic specific sentiment.
And we would also include Twitter analysis. Not just sentiment but what people are talking about.
We are confident that with these improvements we could predict the outcome in the next elections.
Thanks for all the readers. This is the last Baramitter blog. If you are interested in following feedback text analytics business and technology, we will continue the discussion in www.etuma.com.